Neural Information Processing

Recently, recommendation algorithms have been widely used to improve the benefit of businesses and the satisfaction of users in many online platforms. However, most of the existing algorithms generate intermediate output when predicting ratings and the error of intermediate output will be propagated to the final results. Besides, since most algorithms predict all the unrated items, some predicted ratings may be unreliable and useless which will lower the efficiency and effectiveness of recommendation. To this end, we propose a Low-rank and Sparse Matrix Completion (LSMC) method which recovers rating matrix directly to improve the quality of rating prediction. Following the common methodology, we assume the structure of the predicted rating matrix is low-rank since rating is just connected with some factors of user and item. However, different from the existing methods, we assume the matrix is sparse so some unreliable predictions will be removed and important results will be retained. Besides, a slack variable will be used to prevent overfitting and weaken the influence of noisy data. Extensive experiments on four real-world datasets have been conducted to verify that the proposed method outperforms the state-of-the-art recommendation algorithms.

Extreme learning machine (ELM) is a promising learning method for training “generalized” single hidden layer feedforward neural networks (SLFNs), which has attracted significant interest recently for its fast learning speed, good generalization ability and ease of implementation. However, due to its manually selected network parameters (e.g., the input weights and hidden biases), the performance of ELM may be easily deteriorated. In this paper, we propose a novel pre-trained extreme learning machine (P-ELM for short) for classification problems. In P-ELM, the superior network parameters are pre-trained by an ELM-based autoencoder (ELM-AE) and embedded with the underlying data information, which can improve the performance of the proposed method. Experiments and comparisons on face image recognition and handwritten image annotation applications demonstrate that P-ELM is promising and achieves superior results compared to the original ELM algorithm and other ELM-based algorithms.

Community search aims at finding a meaningful community that contains the query node and also maximizes (minimizes) a goodness metric, which has attracted a lot of attention in recent years. However, most of existing metric-based algorithms either tend to include the irrelevant subgraphs in the identified community or have computational bottleneck. Contrary to the user-defined metric algorithm, how can we search the natural community that the query node belongs to? In this paper, we propose a novel community search algorithm based on the concept of k-hop and local distance dynamics model, which can natural capture a community that contains the query node. Extensive experiments on large real-world networks with ground-truth demonstrate the effectiveness and efficiency of our community search algorithm and has good performance compared to state-of-the-art algorithm.

Subway traffic prediction is of great significance for scheduling and anomalies detection. A novel model of multi-scale mixture feedback wavelet neural network(MMFWNN) is proposed to predict the short-term entrance flow of Shanghai subway stations. Firstly, passengers are classified into two categories of commuter and non-commuter by mining the travel pattern and identifying the travel pattern stability, which finds that the non-commuters travel is more susceptible to the meteorology status. The proposed prediction model adds a transitional layer to adapt the feedback mechanism, thus to improve the robustness with associative memorizing and optimization calculation. Thus MMFWNN is advantageous to the nonlinear time-varying short-term traffic flow prediction. We evaluate our model in the Shanghai subway system. The experimental results show that the MMFWNN model is more accurate in predicting the short-term passenger entrance flow in subway stations.

The rapid development of Location-based Social Networks (LBSNs) has led to the great demand of personalized Point-of-interests (POIs) recommendation. Although previous researches have presented a variety of methods to recommend POIs by utilizing social relation, geographical mobility data and user content profile, they fail to address user/location’s cold-start problem with high-dimensional sparse data, and overlook the compatibility of social relation, content based methodology and collaborative filtering. To cope with these challenges, we analyze user’s check-in preference and find that it may be influenced in two spaces, namely Social Propagation Influence Space and Individual Attribute Influence Space. To this end, we propose a Social and Content based Collaborative Filtering Model (SCCF), which consists of a Social Relation Preference based Model (SRPB) considering social friends’ preference and a User Location Content-based Model (ULCB) matching the user attributes with location features. Extensive experiments on real-world datasets firmly demonstrate that the proposed SCCF model outperforms the state-of-the-art approaches while addressing cold-start problems in POI recommendation.

As email workloads keep rising, email servers need to handle this explosive growth while offering good quality of service to users. In this work, we focus on modeling the workload of the email servers of four universities (2 from Greece, 1 from the UK, 1 from Australia). We model all types of email traffic, including user and system emails, as well as spam. We initially tested some of the most popular distributions for workload characterization and used statistical tests to evaluate our findings. The significant differences in the prediction accuracy results for the four datasets led us to investigate the use of a Recurrent Neural Network (RNN) as time series modeling to model the server workload, which is a first for such a problem. Our results show that the use of RNN modeling leads in most cases to high modeling accuracy for all four campus email traffic datasets.

A hybrid sampling technique is proposed by combining Fuzzy C-Mean Clustering and Synthetic Minority Oversampling Technique (FCMSMT) for tackling the imbalanced multiclass classification problem. The mean number of classes is used as the number of instances for applying undersampling and oversampling. Using the mean as the fixed number of the required instances for each class can prevent the within-class imbalance data from being eliminated erroneously during undersampling. This technique can decrease both within-class and between-class errors, and thus can increase the classification performance. The study was conducted using eight benchmark datasets from KEEL and UCI repositories and the results were compared against three major classifiers based on G-mean and AUC measurements. The results reveal that the proposed technique could handle most of the multiclass imbalanced datasets used in the experiments for all classifiers and retain the integrity of the original data.

Visualization methods are important to describe the underlying structure of a data set. When the data is not described as a vector of numerical values, a visualization can be obtained through the reordering of the corresponding similarity matrix. Although several methods of reordering exist, they all need the complete similarity matrix in memory. However, this is not possible for the analysis of dynamic data sets. The goal of this paper is to propose an original algorithm for the incremental reordering of a similarity matrix adapted to dynamic data sets. The proposed method is compared with state-of-the-art algorithms for static data-sets and applied to a dynamic data-set in order to demonstrate its efficiency.

Power consumption is an important part of energy consumption in hydrocracking, which occupies about 43%–47% of the total energy consumption. In the daily production management, the real-time power consumption is manually recorded from the voltmeter. However, it is difficult to collect the power consumption especially in the dynamic adjustment. In this paper, a power consumption prediction model is proposed for dynamic adjustment in the hydrocracking process, which is based on state transition algorithm (STA) and support vector machine (SVM). A SVM regression model is developed to map the complex nonlinear relationship between power parameters and the power consumption in the dynamic adjustment of hydrocracking, and the state transition algorithm is used to optimize the parameters of SVM regression model. The experimental results demonstrate that the prediction accuracy of the model is close to the fitting accuracy and the modeling time is reduced.

The objective of Multi-Task Learning (MTL) is to boost learning performance by simultaneously learning multiple relevant tasks. Identifying and modeling the task relationship is essential for multi-task learning. Most previous works assume that related tasks have common shared structure. However, this assumption is too restrictive. In some real-world applications, relevant tasks are partially sharing knowledge at the feature level. In other words, the relevant features of related tasks can partially overlap. In this paper, we propose a new MTL approach to exploit this partial relationship of tasks, which is able to selectively exploit shared information across the tasks while produce a task-specific sparse pattern for each task. Therefore, this increased flexibility is able to model the complex structure among tasks. An efficient alternating optimization has been developed to optimize the model. We perform experimental studies on real world data and the results demonstrate that the proposed method significantly improves learning performance by simultaneously exploiting the partial relationship across tasks at the feature level.

In this paper, a persona and users’ segmentation model are established by analyzing the power users’ data. In order to further complete the historical database, the paper adopts the method of questionnaire to collect information. Then according to the characteristics of power users, the index system is established, and the index is selected. Different construction methods are adopted for different models. Here, the K-means algorithm is used to cluster the second level indicators in the users’ behavior attribute, and the users’ label is extracted according to the clustering results. Finally, power users’ persona is implemented. It can be proved that the model is effective in dealing with massive data, and provides reliable data support for decision making.

Factorised fuzzy c-means (F-FCM) based on semi nonnegative matrix factorization is a new approach for fuzzy clustering. It does not need the weighting exponent parameter compared with traditional fuzzy c-means, and not sensitive to initial conditions. However, F-FCM does not propose an efficient method to solve the constrained problem, and just suggests to use a lsqlin() function in MATLAB which lead to slow convergence rate and nonconvergence. In this paper, we propose a method to accelerate the convergence rate of F-FCM combining with a non-monotone accelerate proximal gradient (nmAPG) method. We also propose an efficient method to solve the proximal mapping problem when implementing nmAPG. Finally, the experiment results on synthetic and real-world datasets show the performances and feasibility of our method.

It attracts a lot of attention that how to use mobile phone base station data to predict user behavior and design the public traffic route. In this paper, we extend the classic algorithms to design the shuttle bus route. The contribution of this paper is mainly manifested on (1) we integrate the classical machine learning methods DBSCAN and GMM to complete mobile phone base station data modeling, so that to learn the residents’ spatial travel pattern and temporal habits; (2) we apply the Public Route Scale Estimation Model to design the shuttle bus routes and departure intervals based on the modeling results of (1). Experimental results show that our model based on DBSCAN and GMM can effectively mine the significance of historical data of mobile phone base station and can successfully be applied to real-world problems like public traffic route design.

In recent years, there have been many works focusing on combing ratings and reviews to improve the performance of recommender system. Comparing with the rating based algorithms, these methods can be used to alleviate the data sparsity problem in a certain extent. However, they lack the ability to extract the deep semantic information from plaintext reviews. In addition, they do not take the consistence of the latent semantic space of user profiles and item representations into account. To address these problems, we propose a novel method named as Deep Semantic Hybrid Recommendation Method (DSHRM). We utilize deep learning technologies to extract user profiles and item representations from reviews and make sure both of them are in a consistent latent semantic space. We combine ratings and reviews to generate better recommendations. Extensive experiments on real-world datasets show that our method significantly outperforms other six state-of-the-art methods, including LFM, SVD++, CTR, RMR, BoWLF and LMLF methods.

The feature selection is an important step to improve the performance of classifier through reducing the dimension of the dataset, so the time complexity and space complexity are reduced. There are several feature selection methods are used the swarm techniques to determine the suitable subset of features. The sine cosine algorithm (SCA) is one of the recent swarm techniques that used as global optimization method to solve the feature selection, however, it can be getting stuck in local optima. In order to solve this problem, the differential evolution operators are used as local search method which helps the SCA to skip the local point. The proposed method is compared with other three algorithms to select the subset of features used eight UCI datasets. The experiments results showed that the proposed method provided better results than other methods in terms of performance measures and statistical test.

The feature selection (FS) is an important step for data analysis. FS is used to reduce the dimension of data by selecting the relevant features; while removing the redundant, noisy and irrelevant features that lead to degradation of the performance. Several swarm techniques are used to solve the FS problem and these methods provide results better than classical approaches. However, most of these techniques have limitations such as slow convergence and time complexity. These limitations occur due that all the agents update their position according to the best one. However, this best agent may be not the optimal global solution for FS, therefore, the swarm getting stuck in a local solution. This paper proposes an improved Runner-Root Algorithm (RRA). The RRA is combined with chaotic Singer map and opposition-based learning to increase its accuracy. The experiments are performed in eight datasets and the performance of the proposed method is compared against swarm algorithms.

The last decade has witnessed a rapid development of the ensemble clustering technique. Despite the great progress that has been made, there are still some challenging problems in the ensemble clustering research. In this paper, we aim to address two of the challenging problems in ensemble clustering, that is, the local weighting problem and the scalability problem. Specifically, a locally weighted meta-clustering (LWMC) algorithm is proposed, which is featured by two main advantages. First, it is highly efficient, due to its ability of working and voting on clusters. Second, it incorporates a locally weighted voting strategy in the meta-clustering process, which can exploit the diversity of clusters by means of local uncertainty estimation and ensemble-driven cluster validity. Experiments on eight real-world datasets demonstrate the superiority of the proposed algorithm in both clustering quality and efficiency.

Social networks act as the communication channels for people to share various information online. However, spammers who generate spam information reduce the satisfaction of common users. Numerous notable studies have been done to detect social spammers, and these methods can be categorized into three types: unsupervised, supervised and semi-supervised methods. While the performance of supervised and semi-supervised methods is superior in terms of detection accuracy, these methods usually suffer from the dilemma of imbalanced data since the labeled normal users are far more than spammers in real situations. To address the problem, we propose a novel method only relying on normal users to detect spammers. Firstly, a classifier is built from a part of normal and unlabeled samples to pick out reliable spammers from unlabeled samples. Secondly, our well-trained detector, which is based on the given normal users and predicted spammers, can distinguish between normal users and spammers. Experiments conducted on real-world datasets show that the proposed method is competitive with supervised methods.

Computer Science Curricula 2013 (CS2013) is a widely-used standard curricula of computer science, which has been developed jointly by the ACM and the IEEE Computer Society. CS2013 consists of 18 Knowledge Areas (KAs) such as Programming Languages and Software Engineering. Though it is obvious that there are strong interconnections among the KAs, it was hard to investigate the interconnections objectively and quantitatively. In this paper, the interconnections among the KAs of CS2013 are investigated by a data science approach. For this purpose, a collection of actual syllabi from the world’s top-ranked universities was constructed. Then, every actual syllabus is projected to the KA space by a probabilistic model-based method named simplified, supervised Latent Dirichlet Allocation (denoted by ssLDA). Consequently, the following interesting properties of the interconnections among the KAs were discovered: (1) There are the high interconnections among the KAs in each syllabi; (2) A plausible hierarchical structure of the KAs is found by utilizing the interconnections; (3) The structure shows that the KAs are classified into the three principal independent factors (HUMAN, THEORY, and IMPLEMENTATION). The factor of IMPLEMENTATION can be divided into PROGRAMMING and SYSTEM. The factor of SYSTEM can be divided further into DEVICES and NETWORK.

One of the most efficient methods in collaborative filtering is matrix factorization, which finds the latent vector representations of users and items based on the ratings of users to items. However, a matrix factorization based algorithm suffers from the cold-start problem: it cannot find latent vectors for items to which previous ratings are not available. This paper utilizes click data, which can be collected in abundance, to address the cold-start problem. We propose a probabilistic item embedding model that learns item representations from click data, and a model named EMB-MF, that connects it with a probabilistic matrix factorization for rating prediction. The experiments on three real-world datasets demonstrate that the proposed model is not only effective in recommending items with no previous ratings, but also outperforms competing methods, especially when the data is very sparse.

Sentiment classification is a useful approach to analyse the emotional polarity of user reviews, and method based on machine learning has achieved a great success. In the era of Web2.0, the emotional intensity of terms will change with time and events, while a large number of Out-Of-Vocabulary (OOV) terms are appearing. But the method of machine learning pays little attention to them because they focus to reduce the computational complexity. To address the problem, we proposed a dynamic forest model, which can describe the emotional intensity of the term in character granularity, and can append OOV dynamically and adjust their emotional intensity value. Experiments show that in the Chinese environment, our model greatly boosts the performance compared with the method based machine learning, while the time is saved by halves.

Compared to conventional methods, recurrent neural networks and corresponding variants have been proved to be more effective in relation extraction tasks. In this paper, we propose a model that combines a bidirectional long short-term memory network with a multi-attention mechanism for relation extraction. We designed a bidirectional attention mechanism to extract word-level features from a single sentence and chose a sentence-level attention mechanism to focus on features of a sentence set. Our experiments were conducted on a public dataset to evaluate the performance of the model. The experimental results demonstrate that the multi-attention mechanism can make full use of all informative features of a single sentence and a sentence set and our model achieves state-of-the-art performance.

The medical community question answering system (MCQA) which is a new kind of medical information exchange platform is becoming more and more popular. Due to the number of patients is much more than the doctors, resulting in many patients can not get timely answers to their questions. Similar question recommendation is a common approach to solve this problem. The contributions of this paper are two-fold: (1) we propose a Siamese CNN model which measure correlation between questions and answers. (2) We first apply word2vec to learn the semantic relations between words and then construct a similar question retrieval model with answers. The study above can achieve a good performance in the real MCQA data set. It shows that our method can effectively extract similar questions recommendation list, shorten user’s time to wait for an answer and improve user experience as well.

Estimating output changes by input changes is the main task in causal analysis. In previous work, input and output Self-Organizing Maps (SOMs) were associated when conducting causal analysis of multivariate and nonlinear data. Based on the SOM association, a weight distribution of the output conditional on a given input was obtained over the output map space. Such a weighted SOM pattern of the output changes when the input changes. In order to analyze the pattern change, it is important to measure the difference of the patterns. Many methods have been proposed for measuring the dissimilarity of patterns; however, it is still a major challenge to identify how patterns are different. In this paper, we propose a visual approach for analyzing changes to weighted SOM patterns. This approach extracts features that represent the difference of patterns by change and facilitates overall and detailed comparisons of pattern changes. Ecological data are used to demonstrate the usefulness of our approach and the experimental results show that it visualizes the change information effectively.

Mining interesting knowledge from the massive amount of data gathered in wireless sensor networks is a challenging task. Works reported in literature all-confidence measure based associated sensor patterns can captures association-like co-occurrences and the strong temporal correlations implied by such co-occurrences in the sensor data. However, when the user given all-confidence threshold is low, a huge amount of patterns are generated and mining these patterns may not be space and time efficient. Temporal periodicity of pattern appearance can be regarded as an important criterion for measuring the interestingness of associated patterns in WSNs. Associated sensor patterns that occur after regular intervals is called periodic associated sensor patterns. Even though mining periodic associated sensor patterns from sensor data stream is extremely important in many real-time applications, no such algorithm has been proposed yet. In this paper, we propose a compact tree structure called Periodic Associated Sensor Pattern-tree (PASP-tree) and an efficient mining approach for finding periodic associated sensor patterns (PASPs) from WSNs. Extensive performance analyses show that our technique is time and memory efficient in finding periodic associated sensor patterns.

Online multi-label learning is an efficient classification paradigm in machine learning. However, traditional online multi-label methods often need requesting all class labels of each incoming sample, which is often human cost and time-consuming in labeling classification problem. In order to tackle these problems, in this paper, we present online multi-label passive aggressive active (MLPAA) learning algorithm by combining binary relevance (BR) decomposition strategy with online passive aggressive active (PAA) method. The proposed MLPAA algorithm not only uses the misclassified labels to update the classifier, but also exploits correctly classified examples with low prediction confidence. We perform extensive experimental comparison for our algorithm and the other methods using nine benchmark data sets. The encouraging results of our experiments validate the effectiveness of our proposed method.

This paper presents a new method of predicting taxi passenger demands in the central city areas of Seoul and New York based on the temporal and spatial information on predicted values. For the efficiency of the city’s taxi system, investigating the taxi passenger demands is required mainly in the large scaled cities. From this context, this paper proposes a prediction model of combining the conditional transition distribution and the neighboring information on taxi passenger demands. As a result, the proposed method provides higher prediction performances than other methods of homogeneous prediction models.

This paper presents a new way of predicting timely air pollution measure such as the PM$$_{10}$$10 concentration in Seoul based on a new method of combining the global and local estimation models. In the proposed method, the structure of nonlinear dynamics of generating air pollution data series is analyzed by investigating the attractors in the phase space and this structure is used to build the prediction model. Then, the global estimation model such as the network with Gaussian kernel functions is trained for the air pollution series data. Furthermore, the local estimation model which will recover the errors of the global estimation model using the on-line adaptation method, is also adopted. As a result, the proposed prediction model combining the global and local estimation models provides robust performances of predicting PM$$_{10}$$10 concentrations.

Anomaly detection is an important problem in many applications, ranging from medical informatics to network security. Various distribution-based techniques have been proposed to tackle this issue, which try to learn the probabilistic distribution of conventional behaviors and consider the observations with low densities as anomalies. For categorical observations, multinomial or dirichlet compound multinomial distributions were adopted as effective statistical models for conventional samples. However, when faced with small-scale data set containing multivariate categorical samples, these models will suffer from the curse of dimensionality and fail to capture the statistical properties of conventional behavior, since only a small proportion of possible categorical configurations will exist in the training data. As an effective bayesian non-parametric technique, categorical latent Gaussian process is able to model small-scale categorical data through learning a continuous latent space for multivariate categorical samples with Gaussian process. Therefore, on the basis of categorical latent Gaussian process, we propose an anomaly detection technique for multivariate categorical observations. In our method, categorical latent Gaussian process is adopted to capture the probabilistic distributions of conventional categorical samples. Experimental results on categorical data set show that our method can effectively detect anomalous categorical observations and achieve better detection performance compared with other anomaly detection techniques.

Recommender systems can help to relieve the dilemma called information overload. Collaborative filtering is a primary approach based on collective historical ratings to recommend items to users. One of the most competitive collaborative filtering algorithm is matrix factorization. In this paper, we proposed an alternative method. It aims to make users be spatially close to items they like and be far away from items they dislike, by connecting matrix factorization and distance metric learning. The metric and latent factors are trained simultaneously and then used to generate reliable recommendations. The experiments conducted on the real-world datasets have shown that, compared with methods only based on factorization, our method has advantage in terms of accuracy.

Short-term traffic flow prediction plays an important role in intelligent transportation system. Numerous researchers have paid much attention to it in the past decades. However, the performance of traditional traffic flow prediction methods is not satisfactory, for those methods cannot describe the complicated nonlinearity and uncertainty of the traffic flow precisely. Neural networks were used to deal with the issues, but most of them failed to capture the deep features of traffic flow and be sensitive enough to the time-aware traffic flow data. In this paper, we propose a deep bi-directional long short-term memory (DBL) model by introducing long short-term memory (LSTM) recurrent neural network, residual connections, deeply hierarchical networks and bi-directional traffic flow. The proposed model is able to capture the deep features of traffic flow and take full advantage of time-aware traffic flow data. Additionally, we introduce the DBL model, regression layer and dropout training method into a traffic flow prediction architecture. We evaluate the prediction architecture on the dataset from Caltrans Performance Measurement System (PeMS). The experiment results demonstrate that the proposed model for short-term traffic flow prediction obtains high accuracy and generalizes well compared with other models.

In order to master the law of citrus juice odor components changes during the storing process, electronic nose composed of metal-oxide semiconductor (MOS) sensors array is used to monitor the odor during valencia oranges juice storing process. A self-made electronic nose system and experiment are described in detail, after data preprocessing, extreme learning machine (ELM) is used for analysis on samples. Analysis result indicates that the odor synthesized curve derived from the electronic nose technology can reflect overall trend of odor during valencia oranges juice storing process truly and effectively, and the experimental results prove that the E-nose can correctly distinguish the current stage of the stored valencia oranges juice and the classification accuracy of test data set is 96.29% when ELM is used as the classifier, which shows that the E-nose can be successfully applied to the qualitative analysis of citrus.

The location based services have attracted millions of users to share their locations via check-ins. It is highly important to recommend personalized POIs (Points-Of-Interest) to users in terms of their preference learned from historical data. In current research work, users’ check-in behavior is wildly used to model user’s preference. However, the sparsity of the check-in data makes it difficult to capture users’ preferences accurately. This paper proposes a tag-based integrated diffusion recommender system for location recommendation, considering not only social influence but also venue features. Firstly, we model user location preference by combining the preference extracted from check-ins data and short text tips, where sentiment analysis techniques are used. Furthermore, we collect venue information by merging descriptions and tips and then generate tags of each venue, which are processed using keyword extraction approaches. Then we apply the recommendation algorithm with user’s initial preference and obtain the final integrate diffusion results for each user, recommending top-N venues by descending order. We conduct experiments on Foursquare datasets of two cities, the results on both datasets show that our recommender system can produce better performance, providing more personalized and higher novel recommendations.

With the popularity of mobile phones and mobile applications, it becomes possible to collect large-scale mobility data and do research on human mobility. Among these research, relationship mining from location information is a hot topic which has plenty of applications including marketing applications, social studies and even terrorist discovery. This paper focuses on measuring the relationship strength of user pairs according to their meeting events. A novel method using multiple factors extracted from merged meeting events is proposed for measuring relationship. Firstly, meeting events are merged and each merged meeting event is represented by several features, from which multiple factors can be drawn. Specifically, the duration factor and the diameter factor are proposed for measuring relationship on the basis of merged meeting events. Finally, a model synthesizing multiple factors (including location entropy factor, location personal factor, temporal factor, duration factor and diameter factor) is proposed to quantify the relationship between users in an unsupervised way. Experimental results on three different real datasets demonstrate that our method performs significantly more favorable than existing methods on the effectiveness.

With the continuous development of Internet, social networks are becoming more and more complex, and the research on these complex networks has attracted many researchers’ attention. A large number of community discovery algorithms have emerged, among which the label propagation algorithm is widely used because of its simplicity and efficiency. However, this algorithm has poor stability due to the randomness in the label propagation process. To solve the problem, we propose a reinforcement label propagation algorithm (RLPA) in this paper. In RLPA, a similarity matrix is generated from the historical records of classification, which can be adopted to obtain the final result of community detection. The experimental results show that our algorithm can not only get better performance in accuracy, but also has higher stability.

With the rapid development of network technology, people are communicating with each other through a variety of network access, such as computer, mobile phone, tablet, etc., for the sharing of information and interactive behavior. The flow of information is directional, but this directionality is usually hidden. In recent years, link prediction technology has been developed very rapidly in social network analysis. The active and passive of the relationship, in social network, could be identified via undirected relationship network structure. However, this approach only focuses on the topological structure while ignoring the information shared between individuals, which is not suitable for study in terms of information propagation. To solve this problem, we propose a hybrid approach termed DRHM to recover the information sharing direction in networks. It combines not only topology structure but also node content. Since the algorithm is based on edge structure, it is equally applicable to large-scale data set. The experiment has demonstrated that our algorithm performs well in information propagational network.

Point-of-interest (POI) recommendation that suggests new locations for people to visit is an important application in location-based social networks (LBSNs). Compared with traditional recommendation problems, e.g., movie recommendation, geographical influence is a special feature that plays an important role in recommending POIs. Various methods that incorporate geographical influence into collaborative filtering techniques have recently been proposed for POI recommendation. However, previous geographical models have struggled with a problem of geographically noisy POIs, defined as POIs that follow the geographical influence but do not satisfy users’ preferences. We observe that users in the same geographical region share many POIs, and thus we propose the co-geographical influence to filter geographically noisy POIs. Furthermore, we propose the Geo-Pairwise Ranking Matrix Factorization (Geo-PRMF) model for POI recommendation, which incorporates co-geographical influence into a personalized pairwise preference ranking matrix factorization model. We conduct experiments on two real-life datasets, i.e., Foursquare and Gowalla, and the experimental results reveal that the proposed approach outperforms state-of-the-art models.

In order to predict the velocity in driving cycle, first-stage Markov chain (MC) predictor method is adopted. In the traditional Markov prediction model, only one state transition matrix was used to predict the speed. However it will produce a larger error to use the same matrix for predicting speed in different categories of driving cycles. Random Markov-Chain (RMC) model is adopted to improve the accuracy, but the accuracy is still not enough. In this paper, we propose that the state transition matrices in RMC model are divided into two categories: city and highway. Before the prediction, we use the neural network to choose state transition matrix by judging the kinematic parameters of velocity in driving cycles. The simulation results show that the effect of prediction using the state transition matrix after neural network classification is more accurate than no classification. Therefore, the improved RMC model can increase the accuracy of velocity prediction effectively.

We, per the national standard GB14167-2013 of the People’s Republic of China about strength test for ISOFIX anchorage on vehicle seats and taking a new vehicle seat product as research object with the finite element analysis theory, established the finite element model for ISOFIX anchorage on vehicle seats; obtained the stress and strain nephogram of vehicle seats based on HyperWorks software for forward force test and oblique force test; thus provided reference for structural optimization design by analyzing and forecasting the weak parts of vehicle seats.

Phase-mode transformation (PMT) is a commonly used technique to convert a uniform circular array (UCA) into a virtual uniform linear array (ULA). This method restores the Vandemonde structure of the steering vector and makes it easy to apply many existing beamforming algorithms to UCA. One such method is the famous Minimum Variance Distortionless Response (MVDR) algorithm, in which the array gain is equal to unity in the direction of arrival of the desired signal. However, due to the approximation errors of the PMT and signal steering vector mismatches, the performance of these algorithms degrades. To address these two issues, in this paper we develop a robust recursive updating algorithm based on worst-case performance optimization. We show that the proposed algorithm belongs to the class of the diagonal loading technique and the transformation matrix belongs to a certain ellipsoid set. Using the Lagrange multiplier method, we have also derived closed-form solution to the weight vector. Our robust algorithm has low implementation complexity and makes the mean output array SINR consistently close to the optimal one. Numerical experiments have shown that our method outperforms the MVDR algorithm.

Conventional Knowledge-Based Systems (KBS) have no way of detecting or signalling when their knowledge is insufficient to handle a case. Consequently, these systems may produce an uninformed conclusion when presented with a case beyond their current knowledge (brittleness) which results in the KBS giving incorrect conclusions due to insufficient knowledge or ignorance on a specific case. Prudence Analysis (PA) has been shown to be a viable alternative to brittleness in Ripple Down Rules (RDR) knowledge bases. To date, there have been two approaches to Prudence; attribute-based and structural-based prudence. This paper introduces Integrated Prudence Analysis (IPA), a novel Prudence method formed by combining these methods.

Bayesian reinforcement learning (BRL) is an important approach to reinforcement learning (RL) that takes full advantage of methods from Bayesian inference to incorporate prior information into the learning process when the agent interacts directly with environment without depending on exemplary supervision or complete models of the environment. BRL tackles the problem by expressing prior information in a probabilistic distribution to quantify the uncertainty, and updates these distributions when the evidences are collected. However, the expected total discounted rewards cannot be obtained instantly to maintain these distributions after each transition the agent executes. In this paper, we propose a novel idea to adjust immediate rewards slightly in the process of Bayesian Q-learning updating by introducing a state pool technique which could improve total rewards that accrue over a period of time when this pool resets appropriately. We show experimentally on several fundamental BRL problems that the proposed method can perform substantial improvements over other traditional strategies.

Recently, the community is recognizing to an importance of network vulnerability. Also, through the using this vulnerability, attackers can acquire the information of vulnerable users. Therefore, many researchers have been studying about a countermeasure of network vulnerabillty. In recent, the darknet is a received attention to research for detecting action of attackers. The means of darknet are formed a set of unused IP addresses and no real systems of connect to the darknet. In this paper, we proposed an using darknet for the detecting black IPs. So, it was choosen to classification and analysis through source IP of daily darknet traffic. The proposed method prepared 8,192 destination IP addresses in darknet space and collected the darknet traffic during 1 months. It collected total 277,002,257 in 2016, August. An applied results of the proposed process were seen for an effectiveness of pre-detection for real attacks.

In reinforcement learning (RL), the guided policy search (GPS), a variant of policy search method, can encode the policy directly as well as search for optimal solutions in the policy space. Even though this algorithm is provided with asymptotic local convergence guarantees, it can not work in a online way for conducting tasks in complex environments since it is trained with a batch manner which requires that all of the training samples should be given at the same time. In this paper, we propose an online version for GPS algorithm, which can learn policies incrementally without complete knowledge of initial positions for training. The experiments witness its efficacy on handling sequentially arriving training samples in a peg insertion task.

The growing cyber-threats from botnets compel us to devise proper countermeasures to detect infected hosts in an efficient and timely manner. In this paper, botnet-host identification is approached from a new perspective: by exploring the temporal coincidence in botnet activities visible in the darknet, botnet probing campaigns and botnet hosts can be detected with high accuracy and efficiency. The insights to botnet behavioral characteristics and automated detection results obtained from this study suggest a promising expedient for botnet take-down and host reputation management on the Internet.

Motivated by the great efficiency of dynamic time warping (DTW) for time series similarity measure, a Gaussian DTW (GDTW) kernel has been developed for time series classification. This paper proposes an altered Gaussian DTW (AGDTW) kernel function, which takes into consideration each of warping path between time series. Time series can be mapped into a special kernel space where the homogeneous data gather together and the heterogeneous data separate from each other. Classification results on transformed time series combined with different classifiers demonstrate that the AGDTW kernel is more powerful to represent and classify time series than the Gaussian radius basis function (RBF) and GDTW kernels.

Forecasting events like crimes and terrorist activities is a vital important and challenging problem. Researches in recent years focused on qualitative forecasting of a single type event, such as protests or gun crimes. However, events like crimes usually have complicated correlations with each other, and a single type event forecasting cannot meet actual demands. In reality, a quantitative forecasting is more practical for policy making, decision making and police resources allocating. In this paper, we propose an interweaved time series and an interpretative locally connected Recurrent Neural Network model, which forecasts not only whether an event would happen but also how many it would be by each type. Using open source data from Crimes in Chicago provided by Chicago Police Department, we demonstrate our approach more accurately in forecasting the crime events than the existing methods.

Reservoir computing such as Echo State Network (ESN) and Liquid State Machine (LSM) has been successfully applied in dynamical system modeling. However, there is an antagonistic trade-off between the non-linear mapping capacity and the short-term memory capacity in single-reservoir networks, especially when the input signals contain high non-linearity and short-term dependencies. To address this problem, we propose a novel reservoir computing model called Dual-Reservoir Network (DRN), which connects two reservoirs with an unsupervised encoder such as PCA. Specifically, we allow these two adversarial capacities to be decoupled and enhanced in the dual reservoirs respectively. In our experiments, we first verify DRN’s feasibility on an extended polynomial system, which allows us to control the nonlinearity and short-term dependencies of data. In addition, we demonstrate the effectiveness of DRN on the synthesis and real-world time series predictions.

A factored conditional restricted Boltzmann machine (FCRBM) is an efficient, compact model for multi-class temporal data (e.g. multi-label human motion data). However, since all factors in FCRBM are linked to the labels directly, data generated by the model is heavily dependent on the learned tags. In this paper, we propose a tree-based FCRBM model in which the factors are tree-like connected and only part of the factors are directly connected to the labels. The proposed model can make the newly generated data have a variety of sports styles and achieve a smooth transition between the styles using little or even no labeled data.

Short-term traffic flow is a variable affected by many factors. Thus, it is quite difficult to forecast accurately with only one model. The ARIMA model and the SVMs model have their own advantages in terms of linearity and nonlinearity. Therefore, making full use of the advantages of ARIMA model and SVMs model to predict traffic flow can significantly improve the overall effect. The current hybrid approach does not take full account of the characteristics of the data, which cause the effect of hybrid model is not always good. In this paper, first of all, we will use time series analysis and feature analysis to find the characteristics of data. Then, based on the analysis results, we decided to use the method of piecewise to fit the data and make the final prediction. The experiment shows that the piecewise hybrid model can give better play to the advantages of the two models.

Resource prediction promotes dynamic scheduling and energy saving in cloud computing. However, resource prediction becomes a challenge with the diversity and dynamicity of the cloud environment. Existing methods merely focus on single specific resource and ignore the correlation among resources, resulting in inaccurate predictions. Therefore, we propose a trend-matching resources coupled prediction method (TMRCP) based on incremental learning over data stream, which consists of three algorithms. Firstly, to cope with the diversity of the cloud environment, we propose a Resources Utilization Trend Matching algorithm (RUTM), which defines a new similarity measure for multi-dimensional sequences and takes the correlation among resources into consideration. Secondly, we propose a dynamic prediction window adjustment algorithm that selects appropriate prediction length for different resource utilization trends to overcome the disadvantage of fixed window. Thirdly, in response to the sudden changes, we put forward a mixed synthesis algorithm to improve the robustness of the method. Experiments on Google’s cluster usage trace show that the Mean Absolute Percentage Error of TMRCP is 4.7%, 20% better than the state-of-the-art. In addition, the TMRCP is still accurate in multi-step-ahead prediction.

Nowadays mobile applications (a.k.a. app) are playing unprecedented important roles in our daily life and their research has attracted many scholars. However, traditional research mainly focuses on mining app usage patterns or making app recommendations, little attention is paid to the study of app uninstall behaviors. In this paper, we study the problem of app uninstalls prediction based on a machine learning and time series mining approach. Our approach consists of two steps: (1) feature construction and (2) model training. In the first step we extract features from the dynamic app usage data with a time series mining algorithm. In the second step we train classifiers with the extracted features and use them to predict whether a user will uninstall an app in the near future. We conduct experiments on the data collected from AppChina, a leading Android app marketplace in China. Results show that the features mined from time series data can significantly improve the prediction performance.

Time series forecasting has a wide range of applications in society, industry, market, etc. In this paper, a new time series forecasting method (FCD-MLGRU) is proposed for solving short-term forecasting problem. First we decompose the original time series using Filtering Cycle Decomposition (FCD) proposed in this paper, secondly we train the Gated Recurrent Unit (GRU) Neural Network to forecasting the subseries respectively. In the process of training and forecasting, the multi-time-lag sampling and ensemble forecasting method is adopted, which reduces the dependence on the selection of time lag and enhance the generalization and stability of the model. The comparative experiments on the real data sets and theoretical analysis show that our proposed method performs better than other related methods.

We propose here an extended attention model for sequence-to-sequence recurrent neural networks (RNNs) designed to capture (pseudo-)periods in time series. This extended attention model can be deployed on top of any RNN and is shown to yield state-of-the-art performance for time series forecasting on several univariate and multivariate time series.

The ability to predict the ionosphere activity is of interest for several applications such as satellite telecommunications or Global Navigation Satellite Systems (GNSS). A few studies have proposed models able to predict Total Electron Content (TEC) values of the ionosphere locally over measuring stations, but not worldwide for most of them. We propose a method using Deep Neural Networks (DNN) to predict a sequence of global TEC maps consecutive to an input sequence of past TEC maps, by combining Convolutional Neural Networks (CNNs) with convolutional Long Short-Term Memory (LSTM) networks. The numerical experiments show that the approach provides significant improvement over methods implemented for benchmarking and is competitive with state-of-the-art methods while providing global TEC predictions. The proposed architecture can be adapted to any sequence-to-sequence prediction problem.

While wind is an abundant source of energy, integrating wind power into existing electricity grids is a major challenge due to its inherent variability. The ability to accurately predict future generation output would greatly mitigate this problem and is thus extremely valuable. Numerical Weather Prediction (NWP) techniques have been the basis of many wind prediction approaches, but the use of machine learning techniques is steadily gaining ground. Deep Learning (DL) is a sub-class of machine learning which has been particularly successful and is now the state of the art for a variety of classification and regression problems, notably image processing and natural language processing. In this paper, we demonstrate the use of Recurrent Neural Networks, a type of DL architecture, to extract patterns from the spatio-temporal information collected from neighboring turbines. These are used to generate short term wind energy forecasts which are then benchmarked against various prediction algorithms. The results show significant improvements over forecasts produced using state of the art algorithms.

Although neural networks have been very promising tools for chaotic time series prediction, they lack methodology for uncertainty quantification. Bayesian inference using Markov Chain Mont-Carlo (MCMC) algorithms have been popular for uncertainty quantification for linear and non-linear models. Langevin dynamics refer to a class of MCMC algorithms that incorporate gradients with Gaussian noise in parameter updates. In the case of neural networks, the parameter updates refer to the weights of the network. We apply Langevin dynamics in neural networks for chaotic time series prediction. The results show that the proposed method improves the MCMC random-walk algorithm for majority of the problems considered. In particular, it gave much better performance for the real-world problems that featured noise.

Previously the stronger coupling between soil moisture and precipitation in the land-atmosphere interaction have widely been studied. However, few work discusses the causality between them. In this paper, we use Granger causality (GC) and New causality (NC) to detect the causality between soil of different depth moisture and precipitation. Our results demonstrate that the causality between shallow soil moisture and precipitation is greater than that between deep soil moisture and precipitation. And the results also demonstrate that the NC method is much clearer to reveal the causal influence between soil moisture and precipitation than GC method in the time domain.

Haptic perception is to identify different targets from haptic input. Haptic data have two prominent features: sequentially real-time and temporally correlated, which calls for a fixed-budget and recurrent perception procedure. Based on an efficient-robust spatio-temporal feature representation, we handle the problem with a bounded online-sequential learning framework (MBS-ESN), and incorporates the strength of batch-regularization bootstrapping, bounded recursive reservoir, and momentum-based estimation. Experimental evaluations show that it outperforms the state-of-the-art methods by a large margin on test accuracy; and its training performance is superior to most compared models from aspects of computational complexity and storage efficiency.

The current method of dedicated bus arterial coordination priority is mostly based on the arterial coordination control scheme of social vehicle, which makes the dedicated bus arterial coordination priority has many limitations. This paper compare social vehicle traffic flow data with bus traffic flow data which obtained from survey to determine the weighted proportion between them by using spectral clustering (SC) method. And then design multi-period division program for intersection by using Piecewise Aggregate Approximation (PAA). At last we get new arterial coordination control scheme by using graphic method. This paper selects per capita delays as efficiency indicator to measure intersection traffic efficiency. After VISSIM simulation we find out that the new-control-methods outstanding performance on bus traffic efficiency which can decrease the per capita delays reach with public transit-oriented purposes. abstract environment.

We propose a multi-resolution selective ensemble extreme learning machine (MRSE-ELM) method for time-series prediction with the application to the next-step and next-day electricity consumption prediction. Specifically, at the current time stamp, the preceding time-series data is sampled at different time intervals (i.e. resolutions) to constitute the time windows used for the prediction. The value at each sampled point can be certain statistics calculated from its associated time interval. At each resolution, multiple extreme learning machines (ELMs) with different numbers of hidden neurons are first trained. Then, sequential forward selection and least square regression are used to select an optimal set of trained ELMs to constitute the final ensemble model. The experimental results demonstrate that the proposed MRSE-ELM outperforms the best single ELM model across all resolutions. Compared to three state-of-the-art prediction models, MRSE-ELM shows its superiority on the next-step and next-day electricity consumption prediction tasks.

Previous researchers have made some causality hypotheses: the change of stock index causing volatility of economic data or short-run impact of anticipated unemployment rate on stock price. However, they have not reached a consensus. In this article we apply New Causality (NC) method to investigate the causality between Dow Jones Index and the unemployment rate. The results demonstrate stock market is periodically driven by the unemployment rate during all periods, and the causal direction during one ECP and on-going NECP together is uncertain because there may exist two different causal mechanisms in two periods. In this point of view, we conclude that anticipated unemployment rate change results in Dow Jones Index fluctuation in each period. Our conclusion is consistent with the phenomenon that Dow Jones Index was pushed to historical high level after Donald Trump came into power.

A new category called dynamic time series prediction is introduced to address robust “on the fly” prediction needed in events such as natural disasters. A co-evolutionary multi-task learning algorithm is presented which incorporates features from modular and multi-task learning. The algorithm is used for prediction of tropical cyclone wind-intensity. This addresses the need for a robust and dynamic prediction model during the occurrence of a cyclone. The results show that the method addresses dynamic time series effectively when compared to conventional methods.

Influence maximization, first proposed by Kempe, is the problem of finding seed nodes that maximizes the number of affected nodes. However, not only influenced number, but also influence layer is a crucial element which may play an important role in viral marketing. In this paper, we design a new framework, layer-prioritized influence maximization (LPIM), to address the problem of influence maximization with an emphasis on influence layer. The proposed framework is mainly composed of three parts: (1) graph clustering. (2) key node selection. (3) seed node detecting. We also demonstrate the effective and efficient of our proposed framework by experiments on large collaboration networks and complexity analysis respectively.

In the paper, a traffic signal controller based on network was designed after analyzing the development and actuality of traffic signal controllers. The controller consisted of Server, Network Bus and IP Nodes. Power line carrier communication module was designed as the medium of network communication, which could meet the requirements of traffic control network for network bandwidth and communication distance. The software based on Firework computing paradigm can achieve efficient data processing and improve the coordination and optimization ability of traffic control system. With precise actual control effect, the controller can reduce equipment cost, lower difficulties of upgrading and maintenance, and provide complete data support for collaborative optimization of traffic network.

Social media mining has become one of the most popular research areas in Big Data with the explosion of social networking information from Facebook, Twitter, LinkedIn, Weibo and so on. Understanding and representing the structure of a social network is a key in social media mining. In this paper, we propose the Motif Iteration Model (MIM) to represent the structure of a social network. As the name suggested, the new model is based on iteration of basic network motifs. In order to better show the properties of the model, a heuristic and greedy algorithm called Vertex Reordering and Arranging (VRA) is proposed by studying the adjacency matrix of the three-vertex undirected network motifs. The algorithm is for mapping from the adjacency matrix of a network to a binary image, it shows a new perspective of network structure visualization. In summary, this model provides a useful approach towards building link between images and networks and offers a new way of representing the structure of a social network.

Learning microblog users’ interest has important significance for constructing more precise user profile, and can be useful for some commercial applications such as personalized advertisement, or potential customer analysis. Existing works generally utilize text mining or label propagation methods to solve this problem, which leverage either the user’s publicly available comments or the user’s social links, but not both. As we will show, these learning methods achieve limited precision rates. To address this challenge, we consider the interest inference task as a multi-value classification problem, and solve it using a convolutional neural network architecture. We innovatively present an ego social-attribute network model which integrates the target users’ attributes, social links and their comments, and represent the ego SA network as the input fed to CNN. As a result, we assign each microblog user one or more interest labels (such as “loving sports”), which is different from previous approaches using non-uniform interest keywords (such as “basketball”, “tennis”, etc.). Experimental results on SMP CUP and Zhihu dataset showed that the precision rate of user interest inference reached 77.9% at best.

Over the last few years, the amount of Arab sentiment rich data as appearing on the web has been marked with a rapid surge, owing mainly to the remarkable increase noticed in the number of social media users. In this respect, various companies are now turning to online forums, blogs, and tweets with the aim of getting reviews of their products, as drown from customers. Hence, sentiment analysis turns out to lie at the heart of social media associated research, targeted towards detecting people opinion as embedded within the wide range of texts while attempting to capture their pertaining polarities, whether positive or negative.While research associated with English sentiment analysis has already achieved significant progress and success, a remarkable efforts have been made to extend the focus of interest to cover the Arabic language domain. Indeed, most of the Arabic sentiment analysis systems tend to still rely on costly hand-crafted features, where features representation seems to rest on manual pre-processing procedures for the intended accuracy to be achieved. This is mainly due to the Arabic language morphological complexity, linguistic specificities and lack of the resources. For this purpose, deep learning (DL) techniques for Sentiment Analysis turn out to be very versatile and popular. It is in this context that the present paper can be set, with the major focus of the interest being laid on proposing a novel automated information processing systems based DL. The experiment result show that RNN outperforms DNN in term of precision.

Collective action in the provision of pubic goods is analyzed in the framework of three kinds of public goods dilemmas routinely encountered in real-life situations. We study the evolution of cooperation in structured populations within three PGG models: the traditional public goods game (PGG), complementary public goods game (PPGG) and containable public goods game (TPGG), differing in supplying patterns of public goods. In addition, we extend the combination of dual strategy (cooperation and defection) to a portfolio of multiple strategies. We reveal that, is a fundamental property promoting cooperation in groups of selfish individuals, irrespective of which social dilemma applies. For a parallel comparison, it is found that the system in PGG and PPGG can perform comparatively better than TPGG, which reduces the provision of the public goods. Our study can be helpful in effectively portraying the characteristics of cooperative dilemmas in real social systems.

With the growing popularity of social networking services, network public opinion gradually plays an important role in social life. When a controversial network event happens, what people concern about is which opinion in the contrary opinions will be widely accepted by people and how long this event lasts. To solve this problem, we propose a social evolutionary game model based on Hawk-Dove game to simulate how contrary opinions evolve in social network. The effectiveness of our model is validated by actual data. This model can be used to estimate the potential dominant opinion group and the time length of the controversy. Besides, our simulation reveals some special features of the evolution process and results. This study may be useful for network public opinion supervision and market research.

This paper presents methods of predicting categories of questions posted in community-based question answering (CQA) services using deep learning methods, which are implemented with stacked denoising autoencoders (SdA), as well as deep belief networks (DBN). We compare them with conventional machine learning methods, i.e., multi-layer perceptron (MLP) and support vector machines (SVM). We also compare their performance when using dropout regularization. The experimental results indicate that (1) the proposed methods reach much higher prediction precision than that provided by CQA services, (2) deep learning with dropout has higher prediction precision than the conventional machine learning methods, whether or not the dropout regularization is used, i.e., DBN with dropout reaches the highest precision and SdA with dropout reaches the next highest precision among all the methods in general, and the SdA with dropout in a specific case reaches the highest precision across all experiments, (3) increasing the dimensions of feature vectors representing the questions is an effective measure for improving the prediction precision, (4) prediction precision can be further improved using titles in addition to the actual questions and by improving the quality of the corpus used for training.

The proliferation of location-based social networks, makes it possible to record human mobility using an array of points-of-interest (POIs). Exploring the semantic meanings of POIs can be of great importance to many urban computing applications, e.g., personalized route recommendation and user trajectory clustering. Nonetheless, such information is not always available in practice. This paper aims at predicting the category labels, which will provide a succinct summarization of POIs. In particular, we first propose a Location Category Embedding (LCE) model, which projects user POIs and their associated category labels into the same vector space, and then identify the POIs’ most related category labels according to their similarities. To capture the influence that might affect users’ moving behavior, LCE considers sequential pattern, personal preference, and temporal influence, and further models the connection between the POIs and the three factors. Experimental results on two real-world datasets prove the effectiveness of the proposed method.

Community-based question answering (CQA) such as Stack Overflow and Quora face the challenge of providing unsolved questions with high expertise users to obtain high quality answers, which is called question routing. Many existing methods try to tackle this by learning user model from structure and topic information, which suffer from the sparsity issue of CQA data. In this paper, we propose a novel question routing method from the viewpoint of knowledge graph embedding. We integrate topic representations with network structure into a unified Knowledge Graph Question Routing framework, named as KGQR. The extensive experiments carried out on Stack Overflow data suggest that KGQR outperforms other state-of-the-art methods.

Link prediction is an important aspect of complex network evolution analysis. In the existing link prediction algorithms, the sparseness and scale of the target network have a great influence on the prediction results, and the link prediction algorithm based on local random walk is better in solving this problem. However, the existing local random walk link prediction algorithm simplifiy the definition of random walk process between nodes as symmetrical relationship, and ignore the influence of non-visible factors on the relationship of information diffusion between nodes. In this paper, for the first time, we introduce asymmetry and non-visible relationship of the network to the link prediction problem. Exploiting the unequal diffusion weights in different directions resulted from different degrees, we propose an asymmetric local random walk (ALRW) algorithm. In addition, with non-visible relationship to calculate of the similarity index, we propose a grounded asymmetric local random walk (GALRW) algorithm on the basis of ALRW. Compared with existing advanced link prediction algorithms, thorough experiments on typical datasets show that GALRW achieves better performance in prediction accuracy.

Ciphertext retrieval in cloud computing environments requires both security and retrieval efficiency. This paper proposes a Ciphertext Retrieval based on Cloud Pretreatment (CRBCP) based on cloud preprocessing. The scheme divides the cloud into the file server and the index server. Firstly, the program uploads the ciphertext document set to the file server and index server. A lot of preprocessing work is done in the index server and generate an inverted index table. Then, the program uploads the ciphertext retrieval item to the index server. Term Frequency-Inverse Document Frequency (TF-IDF) is used the to get the weight vector of the ciphertext document and the ciphertext retrieval item. Finally, the index server calculates the similarity and returns the result to the client. The simulation results show that the efficiency of encryption and decryption time in the algorithm is obviously higher than that of DjikGentryHaleviVaikuntanathan (DGHV) and Based Vector space model and Homomorphism ciphertext retrieval scheme (BVH). The overall efficiency of ciphertext retrieval in the program is superior than others. In the protection of user data privacy and security under the premise, CRBCP scheme preprocesses the ciphertext in the index server. This will not only greatly improve the efficiency of ciphertext retrieval and reduce the computational pressure on the client, but also fully embody the concept and advantages of cloud computing.

Influence maximization is the problem of finding k seed nodes in a given network as information sources so that the influence cascade can be maximized. To solve this problem both efficiently and effectively, in this paper we propose LAIM: a linear time algorithm for influence maximization in large-scale social networks. Our LAIM algorithm consists of two parts: (1) influence computation; and (2) seed nodes selection. The first part approximates the influence of any node using its local influence, which can be efficiently computed with an iterative algorithm. The second part selects seed nodes in a greedy manner based on the results of the first part. We theoretically prove that the time and space complexities of our algorithm are proportional to the network size. Experimental results on six real-world datasets show that our approach significantly outperforms other state-of-the-art algorithms in terms of influence spread, running time and memory usage.

Precise thyroid nodule classification is a key issue in endocrine clinic domain, which can enhance a patient’s chance for survival. The reports of type-B ultrasound examination are important data source for thyroid nodule classification, and patients with thyroid nodules normally undergo several periodic ultrasound examinations during the process of diagnosis and treatment. However, most of the existing methods rely on feature engineering of single ultrasound reports and they did not take into consideration the historical records of the patients. In this paper, we propose a Hierarchical Recurrent Neural Network (HRNN) for thyroid nodule classification using historical ultrasound reports. HRNN consists of three layers of Long Short-Term Memory (LSTM) Neural Networks. Each LSTM layer is trained to produce the higher-level representations. We evaluate HRNN on real-world thyroid nodule ultrasound reports. The experiment results show that HRNN outperforms the baseline models with ultrasound reports.

Many predictive techniques have been widely applied in clinical decision making such as predicting occurrence of a disease or diagnosis, evaluating prognosis or outcome of diseases and assisting clinicians to recommend treatment of diseases. However, the conventional predictive models or techniques are still not effective enough in capturing the underlying knowledge because it is incapable of simulating the complexity on feature representation of the medical problem domains. This research reports predictive analytical techniques for stroke using deep learning model applied on heart disease dataset. The atrial fibrillation symptoms in heart patients are a major risk factor of stroke and share common variables to predict stroke. The outcomes of this research are more accurate than medical scoring systems currently in use for warning heart patients if they are likely to develop stroke.

The increase in quantity of spatial proteomics data requires a range of analytical techniques to effectively analyse the data. We provide a method of integrating spatial proteomics data together with protein-protein interaction (PPI) networks to enable the extraction of more information. A strong relationship between spatial proteomics and PPI network data was demonstrated. Then a method of converting the PPI network into vectors using spatial proteomics data was explained which allows the integration of the two datasets. The resulting vectors were tested using machine learning techniques and reasonable predictive accuracy was found.

In genetic epidemiology, epistasis has been the subject of several researchers to understand the underlying causes of complex diseases. Identifying gene-gene and/or gene-environmental interactions are becoming more challenging due to multiple genetic and environmental factors acting together or independently. The limitations of current computational approaches motivated the development of a deep learning method in our recent study. The approach trained a multilayered feedforward neural network to discover interacting genes associated with complex diseases. The models are evaluated under various simulated scenarios and compared with the previous methods. The results showed significant improvements in predicting gene interactions over the traditional machine learning techniques. This study is further extended to maximize the predictive performance of the method by tuning the hyperparameters using Cartesian grid and random grid searching. Several experiments are conducted on real datasets to identify higher-order interacting genes responsible for diseases. The findings demonstrated randomly chosen trials are more efficient than trials chosen by grid search for optimizing hyperparameters. The optimal configuration of hyperparameter values improved the model performance without overfitting. The results illustrate top 30 gene interactions responsible for sporadic breast cancer and hypertension.

To understand the processing mechanism of sensory information in the brain, it is necessary to simulate a huge size of network that is represented by a complicated neuron model imitating actual neurons. However, such a simulation requires a very long computation time, failing to perform computer simulation with a realistic time scale. In order to solve the problem of computation time, we focus on the reduction of computation time by GPGPU, providing an efficient method for simulation of huge number of neurons. In this paper, we develop a computational architecture of GPGPU, by which computation of neurons is performed in parallel. Using this architecture, we show that the GPGPU method significantly reduces the computation time of neural network simulation. We also show that the simulations with single and double float precision give little significant difference in the results, independently of the neuron models used. These results suggest that the GPGPU computation with single float precision could be a most efficient method for simulation of a huge size of neural network.

Proposing a robust and fast real-time medical procedure, operating remotely is always a challenging task, due mainly to the effect of delay and dropping of the speed of networks, on operations. If a further stage of prediction is properly designed on remotely operated systems, many difficulties could be tackled. Hence, in this paper, an accurate predictive model, calculating haptics feedback in percutaneous heart biopsy is investigated. A one-layer Long Short-Term Memory based (LSTM-based) Recurrent Neural Network, which is a natural fit for understanding haptics time series data, is utilised. An offline learning procedure is proposed to build the model, followed by an online procedure to operate on new experiments, remotely fed to the system. Statistical analyses prove that the error variation of the model is significantly narrow, showing the robustness of the model. Moreover, regarding computational costs, it takes 0.7 ms to predict a time step further online, which is quick enough for real-time haptic interaction.

Obstructive sleep apnea-hypopnea syndrome is a respiratory disorder characterized by abnormal breathing patterns during sleep. It causes problems during sleep, including loud snoring and frequent awaking. This study proposes a new approach for the detection of apnea-hypopnea events from the raw signal data of nasal airflow using convolutional neural networks. Convolutional neural networks are a prominent type of deep neural networks known for their ability to automatically learn features from high dimensional data without manual feature engineering. We demonstrate the applicability of this technique on a dataset of 24,480 samples (30 s long) extracted from nasal flow signals of 100 subjects in the MESA sleep study. The performance of the convolutional neural network model is compared with another approach that uses a support vector machine model with statistical features generated from the flow signal. Our results show that the convolutional neural network outperformed the support vector machine approach, achieving accuracy and F1-score of 75%.

With the increasing information sharing and other activities conducted on the World Wide Web, the Web has become the main venue for attackers to make troubles. The effective methods to detect Web attacks are critical and significant to guarantee the Web security. In recent years, many machine learning methods have been applied to detect Web attacks. We present a deep learning method to detect Web attacks by using a specially designed CNN. The method is based on analyzing the HTTP request packets, to which only some preprocessing is needed whereas the tedious feature extraction is done by the CNN itself. The experimental results on dataset HTTP DATASET CSIC 2010 show that the designed CNN has a good performance and the method achieves satisfactory results in detecting Web attacks, having a high detection rate while keeping a low false alarm rate.

Chaotic maps are widely applied in many applications. This paper proposes an integrated chaotic system (ICS) to improve the performance of some representative chaotic maps. ICS conducts cascade and nonlinear combination operations to three seed maps such that it has more complex chaotic behaviors and high security levels. A new image encryption algorithm is also developed using ICS. Simulation results on different types of images and security analysis demonstrate that the proposed approach has satisfactory properties in image encryption.

We propose a novel scheme for Android malware detection. The scheme has two extremely fast phases. First term-frequency simhashing (tf-simhashing) extracts a fixed sized vector for each binary file. The hashing algorithm embeds the frequency of n-grams of bytes into the output vector which can be reshaped into an image representation. In the second phase, we propose a convolutional extreme learning machine (CELM) learns to distinguish between hashes of malicious and clean files as a two class classification task. This scalable scheme is extremely fast in both learning and predicting. The results show that tf-simhashing in an image-shape representation together with CELM provides better performance than three non-parametric models and one state-of-the-art parametric model.

The intrusion detection based on deep learning method has been widely attempted for representation learning. However, in various deep learning models for intrusion detection, there is rarely convolutional neural networks (CNN) model. In this work, we propose a image conversion method of NSL-KDD data. Convolutional neural networks automatically learn the features of graphic NSL-KDD transformation via the proposed graphic conversion technique. We evaluate the performance of the image conversion method by binary class classification experiments with NSL-KDD Test$$^+$$+ and Test$$^{-21}$$-21. Different structures of CNN are testified for comparison. On the two NSL-KDD test datasets, CNN performed better than most standard classifier although the CNN did not improve state of the art completely. Results show that the CNN model is sensitive to image conversion of attack data and our proposed method can be used for intrusion detection.

Malicious attack identification plays an essential role in network security monitoring. Current popular technologies are mainly to select a closely related set of attributes from a packet header for fingerprinting malicious attacks. Those methods are not effective enough because malicious attacks can be disguised as normal applications and we cannot observe their characteristics from only the packer’s header. In this paper, we will employ the attributes generated from the entire TCP communication process to identify malicious attacks. A challenging point of our method is how to choose the right attributes from up to 248 properties of TCP flows for fingerprinting low proportion of malicious attacks. A wide variety of real-world viruses are analyzed as the malicious samples, such as extortion virus WannaCry. The experiment results demonstrate that the proposed method can not only fingerprint the viruses but also can accurately identify the types of virus.

Worldwide, the Long Term Evolution Advanced technology has an unprecedented development and popularization in recent years. With the advantages of mobile communication technology, more and more researchers are focused on the security of mobile communication. Until then, some researches about covert channels over the 4th generation mobile communication technology had been proposed. Cantor Expansion is a permutation to a bijection of natural number, so it can be used as a coding scheme for a covert channel. In this paper, a novel class of covert channel based on Cantor Expansion (for decoding) and its inverse operation (for encoding) is proposed and designed for this mobile network. The description, analyses and evaluation of this covert channel will be present in the main part of this paper. Moreover, the peak value of camouflage capability can reach 1470 kbps. Nevertheless it doesnt affect the bandwidth of overt channel and it is difficult to be detected.

It is well known that products for cyber-attacks such as exploits and malware codes are illegally traded on hidden web services called Dark Web that are not indexed by conventional search engines we usually use. In general, it is not easy to capture the whole picture of trade activities on Dark Web because special browsers and tools are needed to visit such dark market sites and forums. And they usually require us to make a registration and/or to pass a qualification test. However, to understand the trends of cyber-attacks, there is no doubt that Dark Web is one of the useful information sources. In this paper, we try to understand the sales trends of illegal products for cyber-attacks from the largest marketplace called AlphaBay, which is relatively easier to collect information without passing any qualification tests, To monitor business trades on Dark Web, we develop an AI web-contents analyzer, which consists of a Tor crawler to collect the product information and a topic analyzer to capture the trends of what people are interested in and popular products of cyber-attacks. For this purpose, we use a topic model called Latent Dirichlet Allocation (LDA) and we show that the topic analysis would be helpful for predicting new cyber-attacks.

In the past, computational models for machine consciousness have been proposed with varying degrees of challenges for implementation. Affective computing focuses on the development of systems that can simulate, recognize, and process human affects which refer to the experience of feeling or emotion. The affective attributes are important factors for the future of machine consciousness with the rise of technologies that can assist humans and also build trustworthy relationships between humans and artificial systems. In this paper, an affective computational model for machine consciousnesses with a system of management of the major features. Real-world scenarios are presented to further illustrate the functionality of the model and provide a road-map for computational implementation.

Measuring psychological concept self-monitoring (SM) is useful for understanding how people employ impression management strategies in their social interactions. Recently, researchers have attempted to utilize the online user data to measure users’ SM value. However, in earlier researches, self-monitoring individuals’ specific behavioral and psychological characteristics haven’ t been sufficiently considered in the process of features extraction. In this paper, motivated by psychologist Snyder’s SM psychological theories, we propose to extract the behavior character of self-monitoring individuals in social network at the macro-level to measure SM. Besides, some other SM relevant features, situational factors, implicit topic words in status updates and demographics are also extracted. Furthermore, a new SM measuring method is presented by exploiting various kinds of users’ online data. The experimental results on a benchmark dataset show that all these features are effective and our SM measuring method can outperform many baseline methods.

Most previous research on indirect reciprocity was in well-mixed population. Distinguishing the interacting network from learning network provides a chance to study indirect reciprocity in networks. Unlike previous research, we propose a coevolution model of cooperation and complex networks via indirect reciprocity, where an individual can interact globally but update strategy locally. Based on this model, we describe the simulation results of coevolution, including the effects of rewiring mechanism on the evolution of cooperation, and how the evolution of cooperation affects networks restructure. Results show that rewiring mechanism favors the evolution of cooperation and the evolution of cooperation can restructure social networks. To understand and explain the results in detail, we graphically depict the snapshots of coevolution process. These findings facilitate us to further understand the evolution of cooperation and the restructure of complex networks.