Knowledge Science, Engineering and Management

Deletion-based sentence compression is frequently formulated as a constrained optimization problem and solved by integer linear programming (ILP). However, ILP methods searching the best compression given the space of all possible compressions would be intractable when dealing with overly long sentences and too many constraints. Moreover, the hard constraints of ILP would restrict the available solutions. This problem could be even more severe considering parsing errors. As an alternative solution, we formulate this task in a reinforcement learning framework, where hard constraints are used as rewards in a soft manner. The experiment results show that our method achieves competitive performance with a large improvement on the speed.

Biomedical question answering system is an important research topic in biomedical natural language processing. To make full use of the semantic knowledge in SNOMED-CT for clinical medical service, we developed a biomedical question answering system based on SNOMED-CT, which has the following characteristics: (a) this system takes the semantic network in SNOMED-CT as a knowledge base to answer the clinical questions posed by physicians in natural language form, (b) a multi-layer nested structure of question templates is designed to map a template into the different semantic relationships in SNOMED-CT, (c) a template description logic system is designed to define the question templates and tag template elements so as to accurately represent question semantics, and (d) a textual entailment algorithm with semantics is proposed to match the question templates in order to consider both the flexibility and accuracy of the system. The experimental results show that the overall performance of the system has reached a high level, which can give 85% of the correct answer and be used as a biomedical question answering system in a real environment.

The goal of authorship attribution is to assign the controversial texts to the known authors correctly. With the development of social media services, authorship attribution for short texts becomes very necessary. In the earlier works, topic models, such as the Latent Dirichlet Allocation (LDA), have been used to find latent semantic features of authors and achieve better performance on authorship attribution. However, most of them focus on authorship attribution for long texts. In this paper, we propose a novel model named Author-Document Topic Model (ADT) which builds the model for the corpus both at the author level and the document level to figure out the problem of authorship attribution for short texts. Also, we propose a new classification algorithm to calculate the similarity between texts for finding the authors of the anonymous texts. Experimental results on two public datasets validate the effectiveness of our proposed method.

The increasing number of documents is leading to more and more topics nowadays. Understanding the relations between different topics evolved in documents become more important and challenging for users. Although many topic models have been devoted to analyzing topics, the study of topics’ potential relevances is still largely limited by various difficulties. Hence, we introduce WalkToTopics, an unsupervised topic mining and analysis model, for inferring potential relevances between different topics. Relying on an advanced feature learning technique to automatically summarize topic’s neighborhood features, WalkToTopics can reveal latent relations between different topics. Compared to existing approaches, our model is able to predict the relationship between any two individual topics of documents, and it does not require any prior knowledge of the existing topics’ relations and dictionaries. Moreover, WalkToTopics is a general model that also can work on exploring topic clusters or extracting sentiments, and can be applied to potential applications, such as ideas tracking and opinion summarization. Finally, we conducted two studies for common users and experts which both quantitatively and qualitatively demonstrate the effectiveness of WalkToTopics in helping users’ understanding of hidden relevances between topics on social media.

Text classification becomes a hot topic nowadays. In reality, the training data and the test data may come from different distributions, which causes the problem of domain adaptation. In this paper, we study a novel learning problem: Distant Domain Adaptation for Text classification (DDAT). In DDAT, the target domain can be very different from the source domain, where the traditional transfer learning methods do not work well because they assume that the source and target domains are similar. To solve this issue we propose a Selective Domain Adaptation Algorithm (SDAA). SDAA iteratively selects reliable instances from the source and intermediate domain to bridge the source and target domains. Extensive experiments show that SDAA has state-of-the-art classification accuracies on the test datasets.

Sentiment analysis is an effective technique and widely employed to analyze sentiment polarity of reviews and comments on the Internet. A lot of advanced methods have been developed to solve this task. In this paper, we propose an attention aware bidirectional GRU (Bi-GRU) framework to classify the sentiment polarity from the aspects of sentential-sequence modeling and word-feature seizing. It is composed of a pre-attention Bi-GRU to incorporate the complicated interaction between words by sentence modeling, and an attention layer to capture the keywords for sentiment apprehension. Afterward, a post-attention GRU is added to imitate the function of decoder, aiming to extract the predicted features conditioned on the above parts. Experimental study on commonly used datasets has demonstrated the proposed framework’s potential for sentiment classification.

Neural network methods have achieved promising results for document-level sentiment classification. Since the popularity of Web 2.0, a growing number of websites provide users with voting and feedback systems (or called social feedback system). However, most existing sentiment classification models only focus on text information while ignoring the social feedback signals from fellow users, despite the association between voting and review predicting. To address this issue, first, we conduct empirical analysis based on a large-scale review dataset to verify the relevance between the social feedback signals and the review predicting. Afterward, we build a hierarchical attention model to generate sentence-level and document-level representations. Finally, we feed the social feedback information into word level and sentence level attention layers. Extensive experiments demonstrate that our model can significantly outperform several strong baseline methods and social feedback signals can promote the performance of attention model.

Manual extraction of business process models from technical documentation is a time-consuming task. Several approaches to generating such process models have been proposed. We present a proposal of a new method for extracting business process from natural language text through intermediate process model using the spreadsheet-based representation. Such intermediate model is transformed into a valid BPMN process model. Our method is enhanced with semantic analysis of the text, allows for quick check of the transformation result and manual correction during this process. As the obtained BPMN model is structured, it is easier to check its correctness.

Automated essay scoring (AES) attempts to rate essays automatically using machine learning and natural language processing techniques, hoping to dramatically reduce the manual efforts involved. Given a target prompt and a set of essays (for the target prompt) to rate, established AES algorithms are mostly prompt-dependent, thereby heavily relying on labeled essays for the particular target prompt as training data, making the availability and the completeness of the labeled essays essential for an AES model to perform. In aware of this, this paper sets out to investigate the impact of data sparsity on the effectiveness of several state-of-the-art AES models. Specifically, on the publicly available ASAP dataset, the effectiveness of different AES algorithms is compared relative to different levels of data completeness, which are simulated with random sampling. To this end, we show that the classical RankSVM and KNN models are more robust to the data sparsity, compared with the end-to-end deep neural network models, but the latter leads to better performance after being trained on sufficient data.

Most NLP tasks are based on large, well-organized corpus in general domain, while limited work has been done in specific domain due to the lack of qualified corpus and evaluation dataset. However domain-specific applications are widely needed nowadays. In this paper, we propose a fast and inexpensive, model-assisted method to train a high-quality distributional model from scattered, unconstructed web pages, which can capture knowledge from a specific domain. This approach does not require pre-organized corpus and much human help, and hence works on the specific domain which can’t afford the cost of artificially constructed corpus and complex training. We use Word2vec to assist in creating term set and evaluation dataset of embroidery domain. Next, we train a distributional model on filtered search results of term set, and conduct a task-specific tuning via two simple but practical evaluation metrics, word pairs similarity and in-domain terms’ coverage. Furthermore, our much-smaller models outperform the word embedding model trained on a large, general corpus in our task. In this work, we demonstrate the effectiveness of our method and hope it can serve as a reference for researchers who extract high-quality knowledge in specific domains.

We often need to cut out a subset of required entities from existing knowledge graphs or websites, when building a knowledge graph for a certain field. In the area of Traditional Chinese Medicine (TCM), we face the task of screening relevant entities from knowledge bases and websites. In this paper, a three-phase TCM entity filter (TCMEF) is proposed, which can identify TCM related entities with high accuracy only using the texts of very short entity titles instead of analyzing the long document texts. The main part of our method is a Short Text LSTM Classifier (STLC), which learns the text style of TCM terms using stroke and character joint features without word segmentation. In addition, an entity representing a person name, which is severe to be classified by STLC, will be picked out by a Person Name Filter (PNF) and further analyzed by a Rich Text Filter (RTF). The filter uses BaiduBaike and HudongBaike (the two largest Chinese encyclopedia websites) as the main data sources. TCMEF gets an F1 score of 0.9275 in classification, which outperforms general word based short text classification algorithms and is close to a Latent Dirichlet Allocation based model (LDA-SVM) using rich texts.

In this paper, we concentrate on tackling the problems of object detection in very-high-resolution (VHR) remote sensing images. The main challenges of object detection in VHR remote sensing images are: (1) VHR images are usually too large and it will consume too much time when locating objects; (2) high false alarm because background dominate and is complex in VHR images. To address the above challenges, a new method is proposed to build two-stage object detection model. Our proposed method can be divided into two processes: (1) we use twice pruning to get region proposal convolutional neural network which is used to predict region proposals; (2) and we use once pruning to get classification convolutional neural network which is used to analyze the result of the first stage and output the class labels of proposals. The experimental results show that the proposed method has high precision and is significantly faster than the state-of-the-art methods on NWPU VHR-10 remote sensing dataset.

Commercial Light-Field cameras provide spatial and angular information, but its limited resolution becomes an important problem in practical use. Different from the conventional images, Light-Field images contain more information of different views that can be used for super-resolution and it makes super-resolution more credible. In this paper, we propose a interpolation based method for Light-Field image super-resolution by taking advantage of the epipolar plane image (EPI) to transfer angular information into spatial information. Firstly, we propose a color recovery framework for undetermined pixels. This framework contains three parts: we estimate the similar-color-diagonal (SCD) for known pixels, we construct a set of filters corresponding to different SCD to generate colors in order to provide a color selection set for undetermined pixel and we propose a W-shaped operator to select a more credible color for undetermined pixel. Finally we use this framework to interpolate EPI and the interpolated EPIs are used to reconstruct a high-resolution image. Experimental results demonstrate that the proposed method outperforms the state-of-art methods for Light-Field spatial super-resolution.

Traditionally, sketch-based image retrieval is mostly based on human-defined features for similarity calculation and matching. The retrieval results are generally similar in contour and lack complete semantic information of the image. Simultaneously, due to the inherent ambiguity of hand-drawn images, there is “one-to-many” category mapping relationship between hand-drawn and natural images. To accurately improve the fine-grained retrieval results, we first train a SBIR general model. Based on the two-branch full-shared parameters architecture, we innovatively propose a deep full convolutional neural network structure model, which obtains mean average precision (MAP) 0.64 on the Flickr15K dataset. On the basis of the general model, we combine the user history feedback image with the input hand-drawn image as input, and use the transfer learning idea to finetune the distribution of features in vector space so that the neural network can achieve fine-grained image feature learning. This is the first time that we propose to solve the problem of personalization in the field of sketch retrieval by the idea of transfer learning. After the model migration, we can achieve fine-grained image feature learning to meet the personalized needs of the user’s sketches.

Multi-target tracking (MTT) has been a research hotspot in the field of computer vision. The objective is forming the trajectory of multiple targets in a given video. However, the useful detection and tracklet relationship during the tracking process are not fully explored in most current algorithms and it leads to the accumulation of errors. We introduce a novel Detection Group, which includes the detections within a temporal and spatial threshold and then model the relationship between Detection Group(DG) and close tracklets. Although the minimum-cost network flow algorithm has been proven to be a successful strategy for multi-target tracking, but it still has one main drawback: due to the fact that useful corresponding detection and tracklet relationships are not well modeled, the network flow based tracker can only model low-level detection relationship without high-level detection set information. To cope with this problem, we extend the classical minimum-cost network flow algorithm within the tracking-by-detection paradigm by incorporating additional constraints. In our experiment, we achieved encouraging result on the MOT17 benchmark and our result is comparable to the current state of the art trackers.

Video-based person re-identification aims to precisely match video sequences of pedestrian across non-overlapped cameras. Existing methods deal with this task by encoding each frame and aggregating them along time. In order to increase the discriminative ability of video features, we propose an end-to-end framework called Multi-grained Fusion Network (MGFN) which aims to keep both global and local information by combining frame-level representations with different granularities. The final video features are generated by aggregating multi-grained representations on both spatial and temporal. Experiments indicate our method achieves excellent performance on three widely used datasets named PRID-2011, iLIDS-VID, and MARS. Especially on MARS, MGFN surpass state-of-the-art result by $$11.5\%$$.

Although the computer and Internet largely improve the convenience of life, they also result in various problems to our work, such as procrastination. Especially, today’s easy access to Internet makes procrastination more pervasive for many people. However, how to accurately assess user procrastination is a challenging problem. Traditional approaches are mainly based on questionnaires, where a list of questions are often created by experts and presented to users to answer. But these approaches are often inaccurate, costly and time-consuming, and thus can not work well for a large number of ordinary people. In this paper, to the best of our knowledge, we are the first to propose to understand and assess people’s procrastination by mining user’s behavioral log on computer. Specifically, as the user’s behavior log is time-series, we first propose a simple procrastination identification model based on the Markov Chain to assess user procrastination. While the simple model can not directly depict reasons of user procrastination, we extract some features from computer logs, which successfully bridge the gap between user behaviors on computer and psychological theories. Based on the extracted features, we design a more sophisticated model, which can accurately identify user procrastination and reveal factors that may cause user’s procrastination. The revealed factors could be used to further develop programs to mitigate user’s procrastination. To validate the effectiveness of our model, we conduct experiments on a real-world dataset and procrastination questionnaires with 115 volunteers. The results are consistent with psychological findings and validate the effectiveness of the proposed model. We believe this work could provide valuable insights for researchers to further exploring procrastination.

Existing works on outlying aspects mining have been focused on detecting the outlying aspects of a single query object, rather than the outlying aspects of a group of objects. While in many application scenarios, methods that can effectively mine the outlying aspects of a query group are needed. To fill this research gap, this paper extends the outlying aspects mining to the group level, and formalizes the problem of group outlying aspect mining. The Earth Move Distance based algorithm GOAM is then proposed to automatically identify the outlying aspects of the query group. The experiment result shows the capability of the proposed algorithm in identifying the group outlying aspects effectively.

Cross-modal retrieval provides a flexible way to find semantically relevant information across different modalities given a query of one modality. The main challenge is to measure the similarity between different modalities of data. Generally, different modalities contain unequal amount of information when describing the same semantics. For example, textual descriptions often contain more background information that cannot be conveyed by images and vice versa. Existing works mostly map the global data features from different modalities to a common semantic space to measure their similarity, which ignore their imbalanced and complementary relationships. In this paper, we propose stacked co-attention networks (SCANet) to progressively learn the mutually attended features of different modalities and leverage these fine-grained correlations to enhance cross-modal retrieval performance. SCANet adopts a dual-path end-to-end framework to jointly learn the multimodal representations, stacked co-attention, and similarity metric. Experiment results on three widely-used benchmark datasets verify that SCANet outperforms state-of-the-art methods, with 19% improvements on MAP in average for the best case.

Data reduction has become one of essential techniques in current knowledge discovery scenarios, dominated by noisy data. The manifold-preserving graph reduction (MPGR) algorithm has been proposed, which has the advantages of eliminating the influence of outliers and noisy and simultaneously accelerating the evaluation of predictors learned from manifolds. Based on MPGR, this paper utilizes the label information to guide the construction of graph and presents a supervised MPGR (SMPGR) method for classification tasks. In addition, we construct a similarity matrix using kernel tricks and develop the kernelized version for SMPGR. Empirical experiments on several datasets show the efficiency of the proposed algorithms.

In the contemporary era, online reviews have an impact on people of all walks of life while choosing appropriate reviews that satisfied user preferences. Personalized reviews selection that is highly relevant to high coverage concerning matching with micro-reviews is the main problem that is considered in this paper. Toward this end, select a personalized subset of reviews are suggested. However, none of the existing research has taken into consideration the personalization of reviews. We proposed a framework known as PeRView for personalized review selection using micro-reviews. The proposed approach shows that our framework can determine and select the best subset of personalized reviews. Based on metric evaluation approach which considered personalized matching score and subset size.

Aiming to the problem of insufficient consideration to the cumulative error and offset which online Global Positioning System (GPS) trajectory data compression based on motion state change and the key point insufficient evaluation of online GPS trajectory data compression based on the offset calculation, an online compression of GPS trajectory data based on motion state change named Synchronous Euclidean Distance (SED) Limited Thresholds Algorithm (SLTA) was proposed. This algorithm used steering angle value and speed change value to evaluate information of trajectory point. At the same time, SLTA introduced the SED to limit offset of trajectory point. So SLTA could reach better information retention. The experiment results show that the trajectory compression ratio can reach about 50%. Compared with Thresholds Algorithm (TA), the average SED error of SLTA can be negligible. For other trajectory data compression algorithms, SLTA’s average angel error is minimum. SLTA can effectively do online GPS trajectory data compression.

Framing detection has emerged to be an important topic in recent natural language processing research. Although several frameworks have been proposed, little is known about how to detect temporal discriminant frames. This study proposes a framework for discovering temporal discriminant frames, with a focus on identifying emergent frames in news discussions of illegal immigration issue. Built on joint non-negative matrix factorization (NMF), we propose the njNMF algorithm, an improved joint matrix factorization algorithm, to detect the temporal frames. We conducted experiments using the njNMF algorithm to identify emergent frames. The results of our experiments show that framing of illegal immigration changes over time, from human trafficking frames, to more recent economic and criminality frames. These findings suggest the utility of our temporal framing approach and can be used as a framing detection tool for policy researchers to understand the role of news framing in public agenda setting.

As a clustering approach with significant potential, the density peak (DP) clustering algorithm is shown to be adapted to different types of datasets. This algorithm is developed on the basis of a few simple assumptions. While being simple, this algorithm performs well in many experiments. However, we find that local density is not very informative in identifying cluster centers and may be one reason for the influence of density parameter on clustering results. For the purpose of solving this problem and improving the DP algorithm, we study the cluster center identification process of the DP algorithm and find that what distinguishes cluster centers from non-density-peak data is not the great local density, but the role of density peaks. We then propose to describe the role of density peaks based on the local density of subordinates and present a better alternative to the local density criterion. Experiments show that the new criterion is helpful in isolating cluster centers from the other data. By combining this criterion with a new average distance based density kernel, our algorithm performs better than some other commonly used algorithms in experiments on various datasets.

In medical intelligent diagnosis, most of the real-world datasets have the class-imbalance problem and some strong correlation features. In this paper, a novel classification model with hierarchical feature representation is proposed to tackle small and imbalanced biomedicine datasets. The main idea of the proposed method is to integrate extreme learning machine-autoencoder (ELM-AE) into the weighted ELM (W-ELM) model. ELM-AE with norm optimization is utilized to extract more effective information from raw data, thereby forming a hierarchical and compact feature representation. Afterwards, random projections of learned feature results view as inputs of the W-ELM. An adaptive weighting scheme is designed to reduce the misclassified rate of the minority class by assigning a larger weight to minority samples. The classification performance of the proposed method is evaluated on two biomedical datasets from the UCI repository. The experimental results show that the proposed method cannot only effectively solve the class-imbalanced problem with small biomedical datasets, but also obtain a higher and more stable performance than other state-of-the-art classification methods.

The goal of network representation learning is to embed each vertex in a network into a low-dimensional vector space. Existing network representation learning methods can be classified into two categories: homogeneous models that learn the representation of vertexes in a homogeneous information network, and heterogeneous models that learn the representation of vertexes in a heterogeneous information network. In this paper, we study the problem of representation learning of heterogeneous information networks which recently attracts numerous researchers’ attention. Specifically, the existence of multiple types of nodes and links makes this work more challenging. We develop a scalable representation learning models, namely SERL. The SERL method formalizes the way to fuse different semantic paths during the random walk procedure when exploring the neighborhood of corresponding node and then leverages a heterogeneous skip-gram model to perform node embeddings. Extensive experiments show that SERL is able to outperform state-of-the-art learning models in various heterogenous network analysis tasks, such as node classification, similarity search and visualization.

Recommendation systems estimate user’s preference to suggest items that might be interesting for them. Recently, implicit feedback recommendation has been steadily receiving more attention because it can be collected on a larger scale with a much lower cost than explicit feedback. The typical methods for recommendation are not well-designed for implicit feedback recommendation. Some effective methods have been proposed to improve implicit feedback recommendation, but most of them suffer from the problems of data sparsity and usually ignore the missing data in implicit feedback. Recent studies illustrate that social information can help resolve these issues. Towards this end, we propose a joint factorization model under the BPR framework utilizing social information. Remarkable, the experimental results show that our method performs much better than the state-of-the-art approaches and is capable of solving implicit problems, which indicates the importance of incorporating social information in the recommendation process to address the poor prediction accuracy.

In recent years, recommender systems based on heterogeneous information networks (HIN) have gained wide attention. In order to generate more attractive recommendations, weighted heterogeneous information network (WHIN) has been proposed, which attaches attribute values to links. The widely-used similarity measures for HIN may fail to capture the semantics of weighted meta-path. This makes designing a similarity measure specially for WHIN more necessary. In this paper, we propose a semantic path-based similarity measure called WgtSim, which is a generalization of PathSim presented by Sun et al. Furthermore, to demonstrate the capability of WgtSim in capturing semantics, we apply WgtSim to recommender system on WHIN to predict ratings given by users. The experiments on two real datasets show that the recommender system with WgtSim outperforms that with previous measures.

Cross-domain algorithms which aim to transfer knowledge available in the source domains to the target domain are gradually becoming more attractive as an effective approach to help improving quality of recommendations and to alleviate the problems of cold-start and data sparsity in recommendation systems. However, existing works on cross-domain algorithm mostly consider ratings, tags and the text information like reviews, cannot use the sentiments implicated in the reviews efficiently. In this paper, we propose a Sentiment Review Pattern Mapping framework for cross-domain recommendation, called SRPM. The proposed SRPM framework can model the semantic orientation of the reviews of users, and transfer sentiment review pattern of users by using a multi-layer perceptron to capture the nonlinear mapping function across domains. We evaluate and compare our framework on a set of Amazon datasets. Extensive experiments on each cross-domain recommendation scenarios are conducted to prove the high accuracy of our proposed SRPM framework.

In recent times, when the Internet is flooded with information, users get overwhelmed with the large amount of data and need some system to narrow down their choices. Recommender systems provide personalized suggestions to the users, giving them a better experience. Data Filtering methods along with many Computational Intelligence (CI) techniques have been used to build and optimize these systems. Here, we introduce a new Recommender System, based on Fuzzy Gravitational Search Algorithm using Hybrid Data Model (FGSA-HDM). FGSA-HDM uses a nature inspired heuristic technique, Gravitational Search Algorithm (GSA), to learn a user’s preference and optimize weightage given to different features which define the user profile. Also, to incorporate the fuzziness of human nature, these features have been represented by Fuzzy sets. The proposed technique, FGSA-HDM, has shown better results than the previously implemented techniques - Pearson Correlation based Collaborative Filtering (PCF), Fuzzy Collaborative Filtering (FCF), Fuzzy Genetic Algorithm based Collaborative Filtering (FG-CF) and Fuzzy Particle Swarm Optimization based Collaborative Filtering (FPSO-CF).

Bayesian networks (BNs) is a dominate model for representing causal knowledge with uncertainty. Causal discovery with BNs requiring large amount of training data for learning BNs structure. When confronted with small sample scenario the learning task is a big challenge. Transfer learning motivated by the fact that people can intelligently apply knowledge learned previously to solve new problems faster or with better solutions, the paper defines a transferable conditional independence test formula which exploit the knowledge accumulated from data in auxiliary domains to facilitate learning task in the target domain, a BNs inductive transfer algorithm were proposed, which learning the Markov equivalence class of BNs. Empirical experiment was deployed, the results demonstrate the effectiveness of the inductive transfer.

In this paper, we are interested in how to explore and utilize the relationship between network communities and semantic topics in order to find the strong explanatory communities robustly. First, the relationship between communities and topics displays different situations. For example, from the viewpoint of semantic mapping, their relationship can be one-to-one, one-to-many or many-to-one. But from the standpoint of underlying community structures, the relationship can be consistent, partially consistent or completely inconsistent. Second, it will be helpful to not only find communities more precise but also reveal the communities’ semantics that shows the relationship between communities and topics. To better describe this relationship, we introduce the transition probability which is an important concept in Markov chain into a well-designed nonnegative matrix factorization framework. This new transition probability matrix with a suitable prior which plays the role of depicting the relationship between communities and topics can perform well in this task. To illustrate the effectiveness of the proposed new approach, we conduct some experiments on both synthetic and real networks. The results show that our new method is superior to baselines in accuracy. We finally conduct a case study analysis to validate the new method’s strong interpretability to detected communities.

Sum-Product Networks (SPNs) are a new class of deep probabilistic model allowing tractable and exact inference. Recently SPNs have been successfully employed as autoencoder framework in Representation Learning. However, SPNs autoencoding mechanism ignores the model structural duality and train the models separately and independently. In this paper, we propose the Dual-SPNs autoencoding mechanism which design model structure as a dual close loop. This approach training the models simultaneously, and explicitly exploiting their structural duality correlation to guide the training process. As shown in extensive multilabel classification experiments, Dual-SPNs autoencoding mechanism prove highly competitive against the ones employing SPNs autoencoding mechanism and other stacked autoencoder architectures.

Modern clinical databases collect a large amount of time series data of vital signs. In this work, we first extract the general representative signal patterns from physiological signals, such as blood pressure, respiration rate and heart rate, referred to as atomic patterns. By assuming the same disease may share the same styles of atomic patterns and their temporal dependencies, we present a probabilistic framework to recognize diseases from physiological data in the presence of uncertainty. To handle the temporal relationships among atomic patterns, Allen’s interval relations and latent variables originated from Chinese restaurant process are utilized to characterize the unique sets of interval configurations of a disease. We evaluate the proposed framework using MIMIC-III database, and the experimental results show that our approach outperforms other competitive models.

Most real-world social networks are usually dynamic (evolve over time), thus communities are constantly changing in memberships. In this paper, an incremental approach based on the coalition formation game theory to identify communities in dynamic social networks is proposed, where the community evolution is modeled as the problem of transformations of stable coalition structures. The proposed approach adaptively update communities from the previous known structures and the changes of topological structure of a network, rather than re-computing in the snapshots of the network at different time steps, such that the computational cost and processing time can be significantly reduced. Experiments have been conducted to evaluate the effectiveness of the proposed approach.

Considerable amounts of business process event logs can be collected by modern information systems. Process discovery aims to uncover a process model from an event log. Many process discovery approaches have been proposed, however, most of them have difficulties in handling large-scale event logs. Motivated by PageRank, in this paper we propose LogRank, a graph-based ranking model, for event log sampling. Using LogRank, a large-scale event log can be sampled to a smaller size that can be efficiently handled by existing discovery approaches. Moreover, we introduce an approach to measure the quality of a sample log with respect to the original one from a discovery perspective. The proposed sampling approach has been implemented in the open-source process mining toolkit ProM. The experimental analyses with both synthetic and real-life event logs demonstrate that the proposed sampling approach provides an effective solution to improve process discovery efficiency as well as ensuring high quality of the discovered model.

Case-based reasoning has become one of the well-sought approaches that supports the development of personalized medicine. It trains on previous experience in form of resolved cases to provide solution to a new problem. In developing a case-based decision support system using case-based reasoning methodology, it is critical to have a good similarity retrieval model to retrieve the most similar cases to the query case. Various factors, including feature selection and weighting, similarity functions, case representation and knowledge model need to be considered in developing a similarity retrieval model. It is difficult to build a single most reliable similarity retrieval model, as this may differ according to the context of the user, demographic and query case. To address such challenge, the present work presents a case-based decision support system with multi-similarity retrieval models and propose contextual bandits learning algorithm to dynamically choose the most appropriate similarity retrieval model based on the context of the user, query patient and demographic data. The proposed framework is designed for DESIREE project, whose goal is to develop a web-based software ecosystem for the multidisciplinary management of primary breast cancer.

Attack path has obviously changed due to multiple-layer structure and the characteristic of failure cross-layer propagation, which changes from static to dynamic and from single layer to multilayer. Attack path exploration is meaningful for simulating the attacker’s intention and is convenient for the defenders to develop a defense mechanism. In this paper, based on a knowledge of target network (i.e., the state of cyber nodes, power flow, node type, voltage, active power, reactive power and time factor etc.), we firstly propose forward and inverse bi-directional solution model that utilizes thread propagation mechanism in the communication network and failure diffusion mechanism in power grid to explore multiple accessible cross-layer attack paths (CLAPs). Thread propagation mechanism considers system vulnerability, threat propagation, and time factor. Failure diffusion mechanism utilizes power flow to trigger load distribution in order to cause attack targets to fail. Secondly, we describe the concept of cross-layer attack path and classify it as four types: Direct Attack Path (DAP), Threat Propagation Attack Path (TPAP), Failure Diffusion Attack Path (FDAP), and Threat Propagation and Failure Diffusion Attack Path (TPFDAP). Thirdly, we propose an assessment method to evaluate the generation probability of CLAPs. Finally, experimental results show that the CLAP of the smart grid can be accurately identified in time, and the defenders can predict the best possible CLAP according to its generation probability. The CLAPs of the same targets are different at the different times and are easily affected by the state of the cyber layer and the tolerance $$\alpha $$ of the physical layer.

In recent digital evolution years, cyber-terrorist activity is increasingly rising all over the world deploying new methods, using advanced technologies and sophisticated weapons. A potential terrorist attack on a large commercial Port could lead to dramatic losses. This work aims to illustrate methods for recognizing cyber-threats and security weaknesses on the ports’ Critical Infrastructures and explores how these issues can be systematically exploited to harm ports and their vicinity. To this end, we follow an asset-centric approach, which employs knowledge representation techniques, to detect vulnerability chains and possible attack-paths on ports’ assets. Considering the results, a realistic coordinated cyber-attack scenario on the application case of the Cruise Vessel Traffic Service is presented to show how cyber-attacks can be realized by terrorists on commercial ports.

The lymph nodes metastasis in the papillary thyroid microcarcinoma (PTMC) can lead to a recurrence of cancer. We hope to take preventive measures to reduce the recurrence rate of the thyroid cancer. This paper presents a decision tree improved by MS-Apriori for the prognosis of lymph node metastasis (LNM) in patients with PTMC, called MsaDtd (Decision tree Diagnosis based on MS-Apriori). The method converts the original feature space into a more abundant feature space, MS-Apriori is used to generate association rules that consider rare items by multiple supports and fuzzy logic is introduced to map attribute values to different subintervals. Then, we filter the ranked rules which consider positive and negative tuples. We improve accuracy through deleting disturbance rules. At last, we use the decision tree to predict LNM by analyzing the affiliation between the instance and rules. Clinical-pathological data were obtained from the First Hospital of Jilin University. The results show that the proposed MsaDtd achieves better prediction performance than other methods on the prognosis of LNM.

The time series of stock prices are non-stationary and non-linear, making the prediction of future price trends much challenging. Inspired by Convolutional Neural Network (CNN), we make convolution on the time dimension to capture the long-term fluctuation features of stock series. To learn long-term dependencies of stock prices, we combine the time convolution with Long Short-Term Memory (LSTM), and propose a novel deep learning model named Time Convolution Long Short-Term Memory (TC-LSTM) networks. TC-LSTM can obtain the stock longer data dependence and overall change pattern. The experiments on two real market datasets demonstrate that the proposed model outperforms other three baseline models in the mean square error.

Recently, large amounts of information on web pages have been emerging in an endless stream. And numerously papers are published on more than three thousands of journals, especially in the field of technology. It’s almost impossible for the user to search the information one by one. The user has to click a lot of links when he or she wants to get information among the thousands of journals, such as the introduction of the journals, impact factor, ISSN and so on. To solve this problem, it’s necessary to develop an automatic method that filter the information out of deep web automatically. The method in this paper is able to help people quickly get needed information classified and extracted. This paper contains the following work: firstly, the method of machine learning, HMM, is used to extract the journal information from the publisher’s website, which improves the generalization ability of using the heuristic method; then, during the data processing step, content extraction technique is used to improve the performance of Hidden Markov Model; finally, we store the extracted information in a structured way and display it. In the experimental step, three algorithms are tested and compared in the accuracy, recall and F-measure, the results show that HMM with content extraction (C-HMM) has the best performance.

We present MedSim, a novel semantic SIMilarity method based on public well-established bio-MEDical knowledge graphs (KGs) and large-scale corpus, to study the therapeutic substitution of antibiotics. Besides hierarchy and corpus of KGs, MedSim further interprets medicine characteristics by constructing multi-dimensional medicine-specific feature vectors. Dataset of 528 antibiotic pairs scored by doctors is applied for evaluation and MedSim has produced statistically significant improvement over other semantic similarity methods. Furthermore, some promising applications of MedSim in drug substitution and drug abuse prevention are presented in case study.

Chinese Named Entity Recognition (NER), as one of basic natural language processing tasks, is still a tough problem due to Chinese polysemy and complexity. In recent years, most of previous works regard NER as a sequence tagging task, including statistical models and deep learning methods. In this paper, we innovatively consider NER as a sequence transformation task in which the unlabeled sequences (source texts) are converted to labeled sequences (NER labels). In order to model this sequence transformation task, we design a sequence-to-sequence neural network, which combines a Conditional Random Fields (CRF) layer to efficiently use sentence level tag information and the attention mechanism to capture the most important semantic information of the encoded sequence. In experiments, we evaluate different models both on a standard corpus consisting of news data and an unnormalized one consisting of short messages. Experimental results showed that our model outperforms the state-of-the-art methods on recognizing short interdependence entity.

Knowledge graphs usually contain much implicit semantic information, which needs to be further mined through semantic inference. Current algorithms can effectively accomplish such task, however they often require a full re-reasoning even when only a few new triples is added to expand the knowledge graph. In this paper, we propose an incremental reasoning algorithm which can effectively avoid re-reasoning over the entire knowledge graph while keeping the relative completeness of the final deduction results. Key to our approach is the filter algorithms which reduce the scale of data that need to be considered and a delay strategy which limit the number of time-consuming iterations while still preserve relative completeness. Extensive experiments and comprehensive evaluations are conducted and experimental results prove that our methods significantly outperform re-reasoning methods.

In relation classification, previous work focused on either whole sentence or key words, meeting problems when sentence contains noise or key words are extracted falsely. In this paper, we propose coarse and fine-grained networks for relation classification, which combine sentence and key words together to be more robust. Then, we propose a word selection network under shortest dependency path (SDP) supervision to select key words automatically instead of pre-processed key words and attention, which guides word selection network to a better feature space. A novel opposite loss is also proposed by pushing useful information in unselected words back to selected ones. In SemEval-2010 Task 8, results show that under the same features, proposed method outperforms state-of-the-art methods for relation classification.