Skip to main content

Über dieses Buch

The two volumes LNCS 11935 and 11936 constitute the proceedings of the 9th International Conference on Intelligence Science and Big Data Engineering, IScIDE 2019, held in Nanjing, China, in October 2019.

The 84 full papers presented were carefully reviewed and selected from 252 submissions.The papers are organized in two parts: visual data engineering; and big data and machine learning. They cover a large range of topics including information theoretic and Bayesian approaches, probabilistic graphical models, big data analysis, neural networks and neuro-informatics, bioinformatics, computational biology and brain-computer interfaces, as well as advances in fundamental pattern recognition techniques relevant to image processing, computer vision and machine learning.



Analysis of WLAN’s Receiving Signal Strength Indication for Indoor Positioning

The location method based on Received Signal Strength Indication (RSSI) ranging has the advantages of low development cost and simple implementation mechanism, and is the mainstream indoor location method nowadays. However, at present, there is a lack of systematic research on the characteristics of Access Point (AP) received signal strength, which cannot meet the signal strength characteristics required by specific indoor positioning environment. Detailedly, in this study, the indoor distribution characteristics of 2.4 GHz RSSI were analyzed experimentally from three influencing factors including antenna orientation of the receiver, type of wireless network card of the receiver and height difference of the transmitting and receiving antenna. The experimental results show that: (1) The RSSI value measured is the strongest When the antenna of the receiver is vertically oriented to the antenna of the transmitter, while the antenna is vertically backward to the transmitter, it is the opposite. The difference between the strongest signal and the weakest signal is 20%–25% at the same test point. (2) The network card with a large measurement range should be selected for data collection on the premise that the quantization step size is 1 and the smoothness is small. Avoiding using the network card for positioning with the maximum limit of measurement range is proposed also. (3) The path loss index of the ranging model is affected by the height difference of the antenna, resulting in a deviation of 0.1–0.2.

Minmin Lin, Zhisen Wei, Baoxing Chen, Wenjie Zhang, Jingmin Yang

Computational Decomposition of Style for Controllable and Enhanced Style Transfer

Neural style transfer has been demonstrated to be powerful in creating artistic images with help of Convolutional Neural Networks (CNN), but continuously controllable transfer is still a challenging task. This paper provides a computational decomposition of the style into basic factors, which aim to be factorized, interpretable representations of the artistic styles. We propose to decompose the style by not only spectrum based methods including Fast Fourier Transform and Discrete Cosine Transform, but also latent variable models such as Principal Component Analysis, Independent Component Analysis, and so on. Such decomposition induces various ways of controlling the style factors to generate enhanced, diversified styled images. We mix or intervene the style basis from more than one styles so that compound style or new style could be generated to produce styled images. To implement our method, we derive a simple, effective computational module, which can be embedded into state-of-the-art style transfer algorithms. Experiments demonstrate the effectiveness of our method on not only painting style transfer but also other possible applications such as picture-to-sketch problems.

Minchao Li, Shikui Tu, Lei Xu

Laplacian Welsch Regularization for Robust Semi-supervised Dictionary Learning

Semi-supervised dictionary learning aims to find a suitable dictionary by utilizing limited labeled examples and massive unlabeled examples, so that any input can be sparsely reconstructed by the atoms in a proper way. However, existing algorithms will suffer from large reconstruction error due to the presence of outliers. To enhance the robustness of existing methods, this paper introduces an upper-bounded, smooth and nonconvex Welsch loss which is able to constrain the adverse effect brought by outliers. Besides, we adopt the Laplacian regularizer to enforce similar examples to share similar reconstruction coefficients. By combining Laplacian regularizer and Welsch loss into a unified framework, we propose a novel semi-supervised dictionary learning algorithm termed “Laplacian Welsch Regularization” (LWR). To handle the model non-convexity caused by the Welsch loss, we adopt Half-Quadratic (HQ) optimization algorithm to solve the model efficiently. Experimental results on various real-world datasets show that LWR performs robustly to outliers and achieves the top-level results when compared with the existing algorithms.

Jingchen Ke, Chen Gong, Lin Zhao

Non-local MMDenseNet with Cross-Band Features for Audio Source Separation

Audio source separation is an important but challenging problem for many applications due to the only available single channel mixed signal. This work proposes a novel Non-Local Multi-scale Multi-band DenseNet model termed as NLMMDenseNet for audio source separation by jointly exploring the long-term dependencies and recovering the missing information around bands’ borders. Specifically, to well leverage the long-term dependencies among the audio spectrogram, we propose a new non-local model by incorporating the non-local layer into MMDenseNet. It enables the proposed model to capture different audio sources features. Besides, the proposed model can also capture cross-band features, which are used to recover the missing information around bands’ borders. The proposed model outperforms state-of-the-art results on the widely-used MIR-1K and DSD100 datasets by taking advantages of global information and bands’ border information.

Yi Huang

A New Method of Metaphor Recognition for A-is-B Model in Chinese Sentences

Metaphor recognition is the bottleneck of natural language processing, and the metaphor recognition for A-is-B mode is the difficulty of metaphor recognition. Compared with phrase recognition, the metaphor recognition for A-is-B mode is more flexible and difficult. To solve this difficult problem, the paper proposes a feature-based recognition method. First, the metaphor recognition problem for A-is-B model is transformed into a classification problem, then four sets of features of upper and lower position, sentence model, class, and Word2Vec are calculated respectively, and feature sets are constructed by using these four sets of features. The experiment uses the SVM model classifier and the neural network classifier to realize the metaphor recognition for the A-is-B mode. The experimental results show that the method using neural network classifier method has better accuracy and recall rate, 96.7% and 93.1%, respectively, but it takes more time to predict a sentence. According to the analysis of the experimental results of the two classifiers, the improved method achieved good results.

Wei-min Wang, Rong-rong Gu, Shou-fu Fu, Dong-sheng Wang

Layerwise Recurrent Autoencoder for Real-World Traffic Flow Forecasting

Accurate spatio-temporal traffic forecasting is a fundamental task for wide applications in city management, transportation area and financial domain. There are many factors that make this significant task also challenging, like: (1) maze-like road network makes the spatial dependency complex; (2) the relationship between traffic flow and time brings non-linear temporal problem; (3) with the larger road network, the difficulty of flow forecasting grows. The prevalent and state-of-the-art methods have mainly been discussed on datasets covering relatively small districts and short time span. To forecast the traffic flow across a wider area and overcome the mentioned challenges, Layerwise Recurrent Autoencoder (LRA) is designed and proposed, in which a three-layer stacked autoencoder (SAE) architecture is used to obtain temporal traffic correlations in three different time scales and for each output of different time scales, a dedicate neural network is used for prediction. The convolutional neural networks (CNN) model is also employed to extract spatial traffic information within the road map for more accurate prediction. To the best of our knowledge, there is no effective method for traffic flow prediction which concerns traffic of city group and LRA is the first one. The experiment is completed on a large real-world traffic dataset to show the performance of the proposed. In the end, evaluations show that our model outperforms the state-of-the-art baselines by 6%–15%.

Junhui Zhao, Tianqi Zhu, Ruidong Zhao, Peize Zhao

Mining Meta-association Rules for Different Types of Traffic Accidents

Association rule method, as one of mainstream techniques of data mining, can help traffic management departments to identify the key contributing factors and hidden patterns in traffic accidents. However, there are still potential links between different accident attributes that have not been revealed, with poor universality of association rules obtained by current methods. In order to overcome the limitations of current methods, this paper proposes a new framework for mining universal rules over different types of traffic accidents, by accounting for the potential dependencies among varied rules suffered from the original methods, and improving the rule selection algorithm. First, different types of traffic accidents are classified and stored separately. Further, the strong association rules for each database are extracted, and then the frequent index approach is applied to organize a meta-rule set with universal applicability. Eventually, all traffic databases are excavated again with different thresholds to get association rules, and meta-rules are integrated into association rules to obtain the universal association rules in the form of a cell group. The proposed method is tested on real traffic databases of nine districts in Shenzhen, China. The results demonstrate that the improved association rules are more universal and representative than existing methods.

Ziyu Zhao, Weili Zeng, Zhengfeng Xu, Zhao Yang

Reliable Domain Adaptation with Classifiers Competition

Unsupervised domain adaptation (UDA) aims to transfer labeled source domain knowledge to the unlabeled target domain. Previous methods usually solve it by minimizing joint distribution divergence and obtaining the pseudo target labels via source classifier. However, those methods ignore that the source classifier always misclassifies partial target data and the prediction bias seriously deteriorates adaptation performance. It remains an open issue but ubiquitous in UDA, and to alleviate this issue, a Reliable Domain Adaptation (RDA) method is proposed in this paper. Specifically, we propose double task-classifiers and dual domain-specific projections to align those easily misclassified and unreliable target samples into reliable ones in an adversarial manner. In addition, the domain shift of both manifold and category space is reduced in the projection learning step. Extensive experiments on various databases demonstrate the superiority of RDA over state-of-the-art unsupervised domain adaptation methods.

Jingru Fu, Lei Zhang

An End-to-End LSTM-MDN Network for Projectile Trajectory Prediction

Trajectory prediction from radar data is an example of a signal processing problem, which is challenging due to small sample sizes, high noise, non-stationarity, and non-linear. A data-driven LSTM-MDN end-to-end network from incomplete and noisy radar measurements to predict projectile trajectory is investigated in this paper. Traditional prediction algorithm usually uses Kalman Filter (KF) or the like to estimate target’s position and speed, then uses the numerical integral, such as Runge-Kutta, Adams, etc. to extrapolate the launch point or impact point, which mainly relies on the accuracy of dynamic models. A Long Short-Term Memory (LSTM) network is designed to estimate the real position from sampled and noisy radar measurements series, and a Mixture Density Network (MDN) is developed for trajectory extrapolation and projectile launch point prediction. These two subnetworks are integrated into an end-to-end network, which is trained by the radar measurement samples of a projectile and the corresponding ground truth of its launch point. Compared with the traditional methods, amount of experiments show that our proposed method is superior to the traditional model-based methods, and its adaptability to the range of initial launch angle is significantly better than the traditional method.

Li-he Hou, Hua-jun Liu

DeepTF: Accurate Prediction of Transcription Factor Binding Sites by Combining Multi-scale Convolution and Long Short-Term Memory Neural Network

Transcription factor binding site (TFBS), one of the DNA-protein binding sites, plays important roles in understanding regulation of gene expression and drug design. Recently, deep-learning based methods have been widely used in the prediction of TFBS. In this work, we propose a novel deep-learning model, called Combination of Multi-Scale Convolutional Network and Long Short-Term Memory Network (MCNN-LSTM), which utilizes multi-scale convolution for feature processing, and the long short-term memory network to recognize TFBS in DNA sequences. Moreover, we design a new encoding method, called multi-nucleotide one-hot (MNOH), which considers the correlation between nucleotides in adjacent positions, to further improve the prediction performance of TFBS. Stringent cross-validation and independent tests on benchmark datasets demonstrated the efficacy of MNOH and MCNN-LSTM. Based on the proposed methods, we further implement a new TFBS predictor, called DeepTF. The computational experimental results show that our predictor outperformed several existing TFBS predictors.

Xiao-Rong Bao, Yi-Heng Zhu, Dong-Jun Yu

Epileptic Seizure Prediction Based on Convolutional Recurrent Neural Network with Multi-Timescale

Epilepsy is a common disease that is caused by abnormal discharge of neurons in the brain. The most existing methods for seizure prediction rely on multi kinds of features. To discriminate pre-ictal from inter-ictal patterns of EEG signals, a convolutional recurrent neural network with multi-timescale (MT-CRNN) is proposed for seizure prediction. The network model is built to complement the patient-specific seizure prediction approaches. We firstly calculate the correlation coefficients in eight frequency bands from segmented EEG to highlight the key bands among different people. Then CNN is used to extract features and reduce the data dimension, and the output of CNN acts as input of RNN to learn the implicit relationship of the time series. Furthermore, considering that EEG in different time scales reflect neuron activity in distinct scope, we combine three timescale segments of 1 s, 2 s and 3 s. Experiments are done to validate the performance of the proposed model on the dataset of CHB-MIT, and a promising result of 94.8% accuracy, 91.7% sensitivity, and 97.7% specificity are achieved.

Lijuan Duan, Jinze Hou, Yuanhua Qiao, Jun Miao

L2R-QA: An Open-Domain Question Answering Framework

Open-domain question answering has always being a challenging task. It involves information retrieval, natural language processing, machine learning, and so on. In this work, we try to explore some comparable methods in improving the precision of open-domain question answering. In detail, we bring in the topic model in the phase of document retrieval, in the hope of exploiting more hidden semantic information of a document. Also, we incorporate the learning to rank model into the LSTM to train more available features for the ranking of candidate paragraphs. Specifically, we combine the results from both LSTM and learning to rank model, which lead to a more precise understanding of questions, as well as the paragraphs. We conduct an extensive set of experiments to evaluate the efficacy of our proposed framework, which proves to be superior.

Tieke He, Yu Li, Zhipeng Zou, Qing Wu

Attention Relational Network for Few-Shot Learning

Few-shot learning aims to learn a model which can quickly generalize with only a small number of labeled samples per class. The situation we consider is how to use the information of the test set to generate the better prototype representation of the training set. In this paper, based on attention mechanism we propose a flexible and efficient framework for few-shot feature fusion, called Attention Relational Network (ARN) which is a three-branch structure of embedding module, weight module and matching module. Specifically, with attention mechanism, the proposed ARN can model adaptively the constribution weights of sample features from embedding module and then generate the prototype representations by weighted fusion of the sample features. Finally, the matching module identify target sample by calculating the matching scores. We evaluated this method on the MiniImageNet and Omniglot dataset, and the experiment proved that our method is very attractive.

Jia Shuai, JiaMing Chen, Meng Yang

Syntactic Analysis of Power Grid Emergency Pre-plans Based on Transfer Learning

To deal with the emergency pre-plans saved by the power grid dispatch department, so that the dispatcher can quickly retrieve and match similar accidents in the pre-plans, then they can learn from the experience of previous relevant situations, it is necessary to extract the information of the pre-plans and extract its key information. Therefore, deep learning method with strong generalization ability and learning ability and continuous improvement of model can be adopted. However, this method usually requires a large amount of data, but the existing labeling data in the power grid field is limited and the manual method for data labeling is a huge workload. Therefore, in the case of insufficient data, this paper aims to solve how to use deep learning method for effective information extraction? This paper modifies the ULMFiT model and uses it to carry out word vector training, adopting transfer learning method to introduce annotating datasets in the open field and combining with the data in the field of power grid to training model. In this way, the semantic relation of power grid domain is introduced into the syntactic analysis of the pre-plans, and we can further complete the information extraction. Experimental verification is carried out in this paper, the results show that, in the case of insufficient corpus or small amount of annotated data, this method can solve the problem of part of speech analysis errors, it can also improve the accuracy of syntactic analysis, and the experimental verifies the effectiveness of this method.

He Shi, Qun Yang, Bo Wang, Shaohan Liu, Kai Zhou

Improved CTC-Attention Based End-to-End Speech Recognition on Air Traffic Control

Recently, many end-to-end speech recognition systems have been proposed aim to directly transcribes speech to text without any predefined alignments. In this paper, we improved the architecture of joint CTC-attention based encoder-decoder model for Mandarin speech recognition on Air Traffic Control speech recognition task. Our improved system include a Vggblstm based encoder, an attention LSTM based decoder decoded with CTC mechanism and a LSTM based ATC language model. In addition, several tricks are used for effective model training, including L2 regularization, attention smoothing and frame skipping. In this paper, we compare our improved model with other three popular end-to-end systems on ATC corpus. Result shows that our improved CTC-attention model outperforms CTC, attention and original CTC-attention model without any tricks and language model. Taken these tricks together we finally achieve a character error rate (CER) of 13.15% and a sentence error rate (SER) of 33.43% on the ATC dataset. While together with a LSTM language model, CER and SER reach 11.01% and 22.75%, respectively.

Kai Zhou, Qun Yang, XiuSong Sun, ShaoHan Liu, JinJun Lu

Revisit Lmser from a Deep Learning Perspective

Proposed in 1991, Least Mean Square Error Reconstruction for self-organizing network, shortly Lmser, was a further development of the traditional auto-encoder (AE) by folding the architecture with respect to the central coding layer and thus leading to the features of Duality in Connection Weight (DCW) and Duality in Paired Neurons (DPN), as well as jointly supervised and unsupervised learning which is called Duality in Supervision Paradigm (DSP). However, its advantages were only demonstrated in a one-hidden-layer implementation due to the lack of computing resources and big data at that time. In this paper, we revisit Lmser from the perspective of deep learning, develop Lmser network based on multiple fully-connected layers, and confirm several Lmser functions with experiments on image recognition, reconstruction, association recall, and so on. Experiments demonstrate that Lmser indeed works as indicated in the original paper, and it has promising performance in various applications.

Wenjin Huang, Shikui Tu, Lei Xu

A New Network Traffic Identification Base on Deep Factorization Machine

Effective network traffic identification has important significance for network monitoring and management, network planning and user behavior analysis. In order to select and extract the most effective attribute as well as explore the inherent correlation between the attributes of network traffic. We proposed a new network traffic identification method based on deep factorization machine (DeepFM) which can classify and do correlation analysis simultaneously. Specifically, we first embed the feature vector into a joint space using a low-rank matrix, then followed by a factorization machine (FM) which handle the low-order feature crosses and a neural network which can learn the high- order feature crosses, finally the low-order feature crosses and high-order feature crosses are fused and give the classified result. We validate our method on Moore dataset which is widely used in network traffic research. Our results demonstrate that DeepFM model not only have a strong ability of network traffic identification but also can reveal some inherent correlation between the attributes.

Zhenxing Xu, Junyi Zhang, Daoqiang Zhang, Hanyu Wei

3Q: A 3-Layer Semantic Analysis Model for Question Suite Reduction

Question generation and question answering are attracting more and more attention recently. Existing question generation systems produce questions based on the given text. However, there is still a vast gap between these generated questions and their practical usage, which acquires more modification from human beings. In order to alleviate this dilemma, we consider reducing the volume of the question set/suite and extracting a lightweight subset while conserving as many features as possible from the original set. In this paper, we first propose a three-layer semantic analysis model, which ensembles traditional language analysis tools to perform the reduction. Then, a bunch of metrics over semantic contribution is carefully designed to balance distinct features. Finally, we introduce the concept of Grade Level and Information Entropy to evaluate our model from a multi-dimensional manner. We conduct an extensive set of experiments to test our model for question suite reduction. The results demonstrate that it can retain as much diversity as possible compared to the original large set.

Wei Dai, Siyuan Sheni, Tieke Hei

Data Augmentation for Deep Learning of Judgment Documents

With the increasing number of machine learning parameters, the requirements on data quantity are getting higher and higher to train a good model. The choice of methods and the optimization of parameters can improve the model while the quality and quantity of the data determine the upper limit of the model. However, in realistic scenarios, it is quite challenging to get a lot of tag data. Therefore, it is natural to realize data augmentation by transforming the original data. We use three methods for data augmentation on different scales of original data in solving the crime prediction problem based on the description of the cases, and find that the effects of data augmentation are different for different models and different fundamental data quantities.

Ge Yan, Yu Li, Shu Zhang, Zhenyu Chen

An Advanced Least Squares Twin Multi-class Classification Support Vector Machine for Few-Shot Classification

In classification tasks, deep learning methods yield high performance. However, owing to lack of enough annotated data, deep learning methods often underperformed. Therefore, we propose an advance version of least squares twin multi-class classification support vector machine (ALST-KSVC) which leads to low computational complexity and comparable accuracy based on LST-KSVC for few-shot classification. In ALST-KSVC, we modified optimization problems to construct a new “1-versus-1-versus-1” structure, proposed a new decision function, and constructed smaller number of classifiers than our baseline LST-KSVC. We empirically demonstrate that the proposed method has better classification accuracy than LST-KSVC. Especially, ALST-KSVC achieves the state-of-the-art performance on MNIST, USPS, Amazon, Caltech image datasets and Iris, Teaching evaluation, Balance, Wine, Transfusion UCI datasets.

Yu Li, Zhonggeng Liu, Huadong Pan, Jun Yin, Xingming Zhang

LLN-SLAM: A Lightweight Learning Network Semantic SLAM

Semantic SLAM is a hot research subject in the field of computer vision in recent years. The mainstream semantic SLAM method can perform real-time semantic extraction. However, under resource-constrained platforms, the algorithm does not work properly. This paper proposes a lightweight semantic LLN-SLAM method for portable devices. The method extracts the semantic information through the matching of the Object detection and the point cloud segmentation projection. In order to ensure the running speed of the program, lightweight network MobileNet is used in the Object detection and Euclidean distance clustering is applied in the point cloud segmentation. In a typical augmented reality application scenario, there is no rule to avoid the movement of others outside the user in the scene. This brings a big error to the visual positioning. So, semantic information is used to assist the positioning. The algorithm does not extract features on dynamic semantic objects. The experimental results show that the method can run stably on portable devices. And the positioning error caused by the movement of the dynamic object can be effectively corrected while establishing the environmental semantic map.

Xichao Qu, Weiqing Li

Meta-cluster Based Consensus Clustering with Local Weighting and Random Walking

Consensus clustering has in recent years become one of the most popular topics in the clustering research, due to its promising ability in combining multiple weak base clusterings into a strong consensus result. In this paper, we aim to deal with three challenging issues in consensus clustering, i.e., the high-order integration issue, the local reliability issue, and the efficiency issue. Specifically, we present a new consensus clustering approach termed meta-cluster based consensus clustering with local weighting and random walking (MC$$^3$$LR). To ensure the computational efficiency, we use the base clusters as the graph nodes to construct a cluster-wise similarity graph. Then, we perform random walks on the cluster-wise similarity graph to explore its high-order structural information, based on which a new cluster-wise similarity measure is derived. To tackle the local reliability issue, all of the base clusters are assessed and weighted according to the ensemble-driven cluster index (ECI). Finally, a locally weighted meta-clustering process is performed on the newly obtained cluster-wise similarity measure to build the consensus clustering result. Experimental results on multiple datasets have shown the effectiveness and efficiency of the proposed approach.

Nannan He, Dong Huang

Robust Nonnegative Matrix Factorization Based on Cosine Similarity Induced Metric

Nonnegative matrix factorization (NMF) is a low-rank decomposition based image representation method under the nonnegativity constraint. However, a lot of NMF based approaches utilize Frobenius-norm or KL-divergence as the metrics to model the loss functions. These metrics are not dilation-invariant and thus sensitive to the scale-change illuminations. To solve this problem, this paper proposes a novel robust NMF method (CSNMF) using cosine similarity induced metric, which is both rotation-invariant and dilation-invariant. The invariant properties are beneficial to improving the performance of our method. Based on cosine similarity induced metric and auxiliary function technique, the update rules of CSNMF are derived and theoretically shown to be convergent. Finally, we empirically evaluate the performance and convergence of the proposed CSNMF algorithm. Compared with the state-of-the-art NMF-based algorithms on face recognition, experimental results demonstrate that the proposed CSNMF method has superior performance and is more robust to the variation of illumination.

Wen-Sheng Chen, Haitao Chen, Binbin Pan, Bo Chen

Intellectual Property in Colombian Museums: An Application of Machine Learning

The purpose of this research is to answer the following guiding question: how can the behavior of museum networks in Colombia be predicted with respect to the protection of intellectual property (copyright, confidential information and use of patents, domain names, industrial designs, use of trademarks) and the interaction of different types of proximity (geographical, organizational, relational, cognitive, cultural and institutional), based on the use of supervised learning algorithms?Among the main findings are that the best learning algorithms to predict the behavior of networks, considering different target variables are the AdaBoost, the naive Bayes and CN2 rule inducer.

Jenny Paola Lis-Gutiérrez, Álvaro Zerda Sarmiento, Amelec Viloria

Hybrid Matrix Factorization for Multi-view Clustering

Multi-view clustering (MVC) has gained considerable attention recently. In this paper, we present a hybrid matrix factorization (HMF) framework which is a combination of the nonnegative factorization and the symmetric nonnegative matrix factorization for MVC. HMF can uncover linear and nonlinear manifold within multi-view dataset. In addition, HMF also learns weights for each view to characterize the contribution of each view to the final common clustering assignment. The proposed model can be solved by nonnegative least squares. Unlike previous approaches, our approach can obtain the clustering results straightforwardly due to the nonnegative constraints. We conduct experiments on multi-view benchmark datasets to verify the effectiveness of our proposed approach.

Hongbin Yu, Xin Shu

Car Sales Prediction Using Gated Recurrent Units Neural Networks with Reinforcement Learning

In this paper, we propose a novel Gated Recurrent Units neural network with reinforcement learning (GRURL) for car sales forecasting. The car sales time series data usually have a small sample size and appear no periodicity. Many previous time series modeling methods, such as linear regression, cannot effectively obtain the best parameter adjustment strategy when fitting the final prediction values. To cope with this challenge and obtain a higher prediction accuracy, in this paper, we combine the GRU with the reinforcement learning, which can use the reward mechanism to obtain the best parameter adjustment strategy while making a prediction. We carefully investigated a real-world time-series car sales dataset in Yancheng City, Jiangsu Province, and built 140 GRURL models for different car models. Compared with the traditional BP, LSTM, and GRU neural networks, the experimental results show that the proposed GRURL model outperforms these traditional deep neural networks in terms of both prediction accuracy and training cost.

Bowen Zhu, Huailong Dong, Jing Zhang

A Multilayer Sparse Representation of Dynamic Brain Functional Network Based on Hypergraph Theory for ADHD Classification

Nowadays, studies on the brain show that the resting brain is still dynamic, and the dynamics of brain functional connectivity remains to be proven, which is very important for the research and diagnosis of mental disorders. In this paper, we apply the Bayesian Connection Change Point Model (BCCPM) to perform dynamic testing on the brain. A sparse model is used to construct a hypergraph to represent the brain function connectivity network, and then the dictionary obtained by sparse learning is used to further extract the features of brain function network. The experimental results on ADHD data show that the accuracy of the proposed method has been improved. Meanwhile, we find that there are obvious differences in the sparse features values of the brain functional networks between patients and normal controls. In addition, the comparison between the proposed method with/without the BCCPM demonstrated the importance of dynamic detection further.

Yuduo Zhang, Zhichao Lian, Chanying Huang

Stress Wave Tomography of Wood Internal Defects Based on Deep Learning and Contour Constraint Under Sparse Sampling

In order to detect the size and shape of defects inside wood using stress wave technology under sparse sampling, a novel tomography algorithm is proposed in this paper. The method uses instrument to obtain the stress wave velocity data by sensors hanging around the timber equally, visualizes those data, and reconstructs the image of internal defects with estimated velocity distribution. The basis of the algorithm is using deep learning to assist stress wave tomography to resist signal reduction. First, training CNN model with a large number of generated simulation samples and two-level defect location labeling, and detecting the defective region in wood. Second, using CNN detection results to assist tomography algorithm to precisely estimate the defective area with contour constraint including deepening and weakening operations. Both simulation and wood samples were used to evaluate the proposed method. Effect of CNN detection results on tomography and the shape of the imaging results were both analyzed. The comparison results show that the proposed method always can produce high quality reconstructions with clear edges, when the number of sensors is decreased from 12 to 6.

Xiaochen Du, Jiajie Li, Hailin Feng, Heng Hu

Robustness of Network Controllability Against Cascading Failure

Controllability of networks widely existing in real-life systems have been a critical and attractive research subject for both network science and control systems communities. Research in network controllability has mostly focused on the effects of the network structure on its controllability, and some studies have begun to investigate the controllability robustness of complex networks. Cascading failure is common phenomenon in many infrastructure networks, which largely affect normal operation of networks, and sometimes even lead to collapse, resulting in considerable economic losses. The robustness of network controllability against the cascading failure is studied by a linear load-capacity model with a breakdown probability in this paper. The controllability of canonical model networks under different node attack strategies is investigated, random failure and malicious attack. It is shown by numerical simulations that the tolerant parameter of load-capacity model has an important role in the emergence of cascading failure, independent to the types of network. The networks with moderate average degree are more vulnerable to the cascading failure while these with high average degree are very robust. In particular, betweenness attack strategy is more harmful to the network controllability than degree attack one, especially for the scale-free networks.

Lv-lin Hou, Yan-dong Xiao, Liang Lu

Multi-modality Low-Rank Learning Fused First-Order and Second-Order Information for Computer-Aided Diagnosis of Schizophrenia

The brain functional connectivity network (BFCN) based methods for diagnosing brain diseases have shown great advantages. At present, most BFCN construction strategies only calculate the first-order correlation between brain areas, such as the Pearson correlation coefficient method. Although the work of the low-order and high-order BFCN construction methods exists, there is very little work to integrate them, that is, to design a multi-modal BFCN feature selection and classification method to combine low-order and high-order information. This may affect the performance of brain disease diagnosis. To this end, we propose a multi-modality low-rank learning framework jointly learning first-order and second-order BFCN information and apply it to the diagnosis of schizophrenia. The proposed method not only embeds the correlation information of multi-modality data in the learning model, but also encourages the cooperation between the first-order and the second-order BFCN by combining the ideal representation term. The experimental results of the three schizophrenia datasets (totally including 168 patients and 163 normal controls) show that our proposed method achieves promising classification results in the diagnosis of schizophrenia.

Huijie Li, Qi Zhu, Rui Zhang, Daoqiang Zhang

A Joint Bitrate and Buffer Control Scheme for Low-Latency Live Streaming

Live video streaming has experienced explosive growth on the mobile Internet. Unlike on-demand streaming, live video streaming faces more challenges due to the strong requirement of low latency. To balance several inherently conflicting performance metrics and improve the overall quality of experience (QoE), the adaptive bitrate algorithm is widely used under time-varying network conditions. However, it does not perform well at low latency. In this paper, we present a joint bitrate and buffer control scheme (JBBC) for low-latency live streaming based on latency-constrained bitrate adaptation and playback rate adaptation. Experiments demonstrate that the proposed algorithm has better performance on overall QoE than most existing adaptive schemes, achieving a more stable bitrate selection and relatively lower delay on the premise of almost no rebuffering.

Si Chen, Yuan Zhang, Huan Peng, Jinyao Yan

Causal Discovery of Linear Non-Gaussian Acyclic Model with Small Samples

Linear non-Gaussian Acyclic Model (LiNGAM) is a well-known model for causal discovery from observational data. Existing estimation methods are usually based on infinite sample theory and often fail to obtain an ideal result in the small samples. However, it is commonplace to encounter non-Gaussian data with small or medium sample sizes in practice. In this paper, we propose a Minimal Set-based LiNGAM algorithm (MiS-LiNGAM) to address the LiNGAM with small samples. MiS-LiNGAM is a two-phase and greedy search algorithm. Specifically, in the first phase, we find the skeleton of the network using the regression-based conditional independence test, which helps us reduce the complexity in finding the minimal LiNGAM set of the second phase. Further, this independence test we applied guarantees the reliability when the number of conditioning variables increases. In the second phase, we give an efficient method to iteratively select the minimal LiNGAM set with the skeleton and learn the causal network. We also present the corresponding theoretical derivation. The experimental results on simulated networks and real networks are presented to demonstrate the efficacy of our method.

Feng Xie, Ruichu Cai, Yan Zeng, Zhifeng Hao

Accelerate Black-Box Attack with White-Box Prior Knowledge

We propose an efficient adversarial attack method in the black-box setting. Our Multi-model Efficient Query Attack (MEQA) method takes advantage of the prior knowledge on different models’ relationship to guide the construction of black-box adversarial instances. The MEQA method employs several gradients from different white-box attack models and further the “best” one is selected to replace the gradient of black-box model in each step. The gradient composed by different model gradients will lead a significant loss to the black-box model on these adversarial pictures and then cause misclassification. Our key motivation is to estimate the black-box model with several existing white-box models, which can significantly increase the efficiency from the perspectives of both query sampling and calculating. Compared with gradient estimation based black-box adversarial attack methods, our MEQA method reduces the number of queries from 10000 to 40, which greatly accelerates the black-box adversarial attack. Compared with the zero query black-box adversarial attack method, which also called transfer attack method, MEQA boosts the attack success rate by 30%. We evaluate our method on several black-box models and achieve remarkable performance which proves that MEQA can serve as a baseline method for fast and effective black-box adversarial attacks.

Jinghui Cai, Boyang Wang, Xiangfeng Wang, Bo Jin

A Dynamic Model + BFR Algorithm for Streaming Data Sorting

Streaming data is widely generated in our lives. This has promoted a lot of research on streaming data mining, such as streaming data clustering and filtering. In our work, we present a problem about data stream processing, namely, streaming data sorting. There are some important characteristics of streaming data. Firstly, streaming data comes in the form of streams. It is usually assumed that streaming data is infinite, so it cannot be stored completely in memory. Secondly, we must process the streaming data in real time, otherwise we may lose the opportunity to deal with it forever. Based on these characteristics, we propose a dynamic algorithm that can make full use of memory and minimize error to solve the problem of streaming data sorting, which is further combined with the BFR algorithm to sort a particular type of streaming data. Some experiments are conducted to confirm the effectiveness of the proposed algorithms.

Yongwei Tan, Ling Huang, Chang-Dong Wang

Smartphone Behavior Based Electronical Scale Validity Assessment Framework

In the study, we developed a smartphone-based electronical scale validity assessment framework. 374 college students are recruited to fill in Beck Depression Inventory. A total of 544 filling of scales are collected, which may be filled accordingly or concealed. Via an electronical scale based WeChat applet and backend application, temporal and spatial behavioral data of subjects during the scale-filling process are collected. We established an assessment model of the validity of the scale-filling based on the behavior data with machine learning approaches. The result shows that smartphone behavior has significant features in the dimension of time and space under different motivations. The framework achieves an valuable assessment of the effectiveness of the scale, whose key indicators such as accuracy, sensitivity and precision are over 80% under multiple dimension behavior data classification. The framework has a good application prospect in the field of psychological screening.

Minqiang Yang, Jingsheng Tang, Longzhe Tang, Bin Hu

Discrimination Model of QAR High-Severity Events Using Machine Learning

The Quick Access Recorder (QAR) is an airborne equipment designed to store raw flight data, which contains a mass amount of safety related parameters such as flap angle, airspeed, altitude, etc. The assessment of QAR data is of great significance for the safety of civil aviation and the improvement of pilots skills. The existing QAR assessment approaches mainly utilizes the exceedance detection (ED) that relies on the pre-defined parameter threshold, which could miss potential flight risks. In this paper, we perform anomaly detection on the takeoff and landing phases based on an improved random forest (RF) method. The evaluation is performed on the dataset generated by a fleet of B-737NG, which shows that the method is able to discriminate the high-severity events accurately on the high dimensional multivariate time series, which also shows that the model can identify the events with potential risk pattern on the imbalanced dataset even if the event has not been pre-defined before.

Junchen Li, Haigang Zhang, Jinfeng Yang

A New Method of Improving BERT for Text Classification

Text classification is a basic task in natural language processing. Recently, pre-training models such as BERT have achieved outstanding results compared with previous methods. However, BERT fails to take into account local information in the text such as a sentence and a phrase. In this paper, we present a BERT-CNN model for text classification. By adding CNN to the task-specific layers of BERT model, our model can get the information of important fragments in the text. In addition, we input the local representation along with the output of the BERT into the transformer encoder in order to take advantage of the self-attention mechanism and finally get the representation of the whole text through transformer layer. Extensive experiments demonstrate that our model obtains competitive performance against state-of-the-art baselines on four benchmark datasets.

Shaomin Zheng, Meng Yang


Weitere Informationen

Premium Partner