Abstract:The rapid development of information technology has led to information overload.Recommendation is one of 他the most effective ways to solve the information overload.In recent years,the rapid development of deep learning has also led to the advancement of recommender systems,and various deep learning based recommendation algorithms have emerged one after another.However,due to the large number of candidate items and the dynamic evolving of user interests,deep learning based recommendation algorithms suffer from computational burden of online recommendation.Therefore,it is almost impossible for these algorithms to be deployed alone in practice.With the development of deep learning based recommendation,the item recallingtechniques(also called approximated search techniques) has also made significant progress.This paper first introduces the research progress of the item recalling techniques based on the nearest neighbor search,and then discusses the research progress of the item recallingtechniquesbased on the maximum inner product search from the perspectives of indexing,locality sensitive hash,learning to hash and vector quantization.

Abstract:With the rapid development of Internet of things and cloud computing,big data and its related science have become the focus of industry and academia.In this paper,we analyzes big data science from the perspective of paradigm theory and expounds the difference and connection between big data and traditional research.Three major challenges brought by big data are proposed in perspective of machine learning,with the corresponding scientific problems following.In addition,this paper introduces several insights of big data science from the perspective of the forth paradigm and its positive significance.In the end,we summarize and look forward to the challenges of big data science in the future.

Abstract:In recent years,as one of the important issues in natural language processing,the human-machine dialogue system has received more and more attention.Methods of deep learning based on big data are widely used in dialogue systems.In this paper,the background of human-machine dialogue system is firstly introduced,then dialogue system based on multiple-round interaction is taken as an example.This paper emphasizes on task-oriented and non-task-oriented dialogue systems,elaborates the main types and current research progress.After that,this paper provides an overview of some main methods of evaluating the dialogue system.Finally,based on the current research status,the prospect of some future research directions on human-machine dialogue system are discussed.

Abstract:The goal of transfer learning is to solve the problem of insufficient training samples in the target domain.It can transfer the acquired knowledge from related source domain to the target domain.It relaxes two basic assumptions in traditional machine learning:the training samples and the new test samples satisfy the conditions of independent and identical distribution; furthermore,there must be enough training samples to learn a good classification model.According to whether the feature space of the source domain and the target domain are the same,it can be divided into homogeneous transfer learning and heterogeneous transfer learning.This paper mainly reviews the related research progress of homogeneous transfer learning,introduces the theory,algorithm and application of homogeneous transfer learning,and points out the hotissues of homogeneous transfer learning.

Abstract:Since massive news articles are generated and posted online,news recommendation has become an important way to alleviate user information overload and achieve personalized news information access,which has been widely used in many news websites and news APPs to improve user experience.Different from the traditional product recommendation,in the scenario of news recommendation,the news articles are generated very quickly,and the semantic meaning of news articles needs to be captured from the original news textual content,which bring huge challenges to the traditional recommendation methods which are based on IDs and features.In addition,users' news reading interests are highly diverse and dynamic,making it difficult to accurately model users.In this paper we will introduce several deep learning based news recommendation algorithms,and explore several future directions of news recommendation.

Abstract:Sentiment analysis aims to extract users' sentimentsand opinions about or their attitude towarda specific product,service,or event.The lack of labeled data is a significant challenge in sentiment analysisand will deteriorate the performance of the classifierin a supervised sentiment analysis task.The cross-domain approach has been shown to be effective in addressing this problem.However,the inherent difference between the source and target domains will make it difficult for the classifier to be adaptive to the target domain.In this paper,we propose a novel method to use the available labeled data,however few they may be,in the target domain to enhance the domain adaption.Specifically,we present a cross-domain sentiment classification model using the capsule network.Based on this architecture,we design extra capsule layers for domain adaption.Extensive experiments with real-world datasets prove that our proposed model outperforms baselines by a large margin.

Abstract:To solve the problem of the lack of use of shallow image features in image captions and insufficient extraction of image objects,an image caption generation algorithm based on attention image feature extraction is proposed.Through context information of a language model,adaptive attention weight assignment is performed on different depth image features to ensure that the attention-grabbing image features guide the image caption generation,thereby improving the image caption effect.In the MSCOCO test set,the BLEU-1 and CIDEr scores of the proposed algorithm reached 0.752 and 0.934,respectively,thus verifying the effectiveness of the proposed method.

Abstract:To solve the implicit recommendation problems,Bayesian Personalized Ranking(BPR) algorithm has become one the most representative pairwise methods.Generally,BPR assumes that users keep higher preference on observed items than unobserved items.In this paper,we introduce Multi-pair Bayesian Personalized Ranking (MBPR),a novel pairwise method to further investigate the preference about the large number of unobserved feedbacks.First,we propose an enhanced pairwise assumption based on the traditional pairwise assumption adopted by BPR.Specifically,we divide the large unobserved item set into two parts:uncertain item set and possibly negative item set for each user.Based on this,a new multi-pair pairwise objective function is proposed to learn users' preference.To solve the sampling task in MBPR,an adaptive sampling strategy is then proposed to dynamically draw uncertain feedbacks from unobserved item set.Finally,empirical studies show that our algorithms can improve the ranking performance of BPR.

Abstract:The traditional multi-view dictionary learning algorithm is designed to take advantage of the correlation between multi-view data and fails to consider the distinctiveness of the multi-view data,which may reduce the performance of dictionary.Inspired by this observation,we present a multi-view dictionary learning based on the intra-view atom inconsistency algorithm.The algorithm learns class-specific dictionaries and the shared class dictionary for each view and calculates the minimum of the coding coefficient variance to reduce the distinctiveness of inter-view dictionaries.In addition,the minimization of the weighted sum of the distance between the coding coefficients between each view and the mean of coding coefficients for all views restrict the contribution of the corresponding features.Then,we embed the inconsistency constraint into the intra-view dictionaries to reduce redundancy.Finally,two datasets (AR and Extended Yale B datasets) were used to validate the effectiveness of the proposed algorithm.

Abstract:Understanding the key factors that influence student performances will help students,teachers,and administrators to improve the performance of the students.To this end,the density-based global K-means algorithm is adopted to perform cluster analysis of the student performance data from the UCI machine learning repository for two secondary education Portuguese schools and for a senior middle school of the Pucheng county in the Shaanxi province.The results for the two Portuguese schools reveal that student performance is strongly related to the specific school where the student is enrolled,and location of residence,mother's education level,and if the network is available or not in the family.Education level of the father,the time the student takes on the way to school,the willingness of the student to go to college,and whether the student is in love are factors affecting the student performance to some extent.The results of the third senior middle school demonstrate that student performance is strong related to their guardians,parents' age,parents' education level,learning attitude of the student,and the time the student devotes to courses after classes.In addition,the results indicate that scores of a student for the upcoming examination can be predicted with the available ones and that the predicted scores coincide with the actual ones.The studies in this paper demonstrate that student performance is strong related to parents' education level,especially to mother's education level.The higher the level of education of the mother,the better the student performance.Parents cannot ignore their role in the individual growth of children.It is important to teach students to study actively to improve their achievements.Finally,it is imperative that the education gap between the urban and rural areas is narrowed.

Abstract:Regional economic development not only impacts regional politicization and economic construction but also improves national comprehensive competitiveness.How to predict the relationship between regional stability and economic development is an important problem.Accurately and quantitatively explaining development trends by using historical regional economic development data to analyze future development of the region can be difficult owing to the complexity of economic development.The goal of this work is to explore the dynamic,complex,and interactive process of the conflict caused by environmental change and analyze the potential correlation between environmental change and economic growth by regarding regional stability as a measure of the relationship between environmental change and economic growth.The main aspects of this work are as follows:1)Using the Fitness and Complexity algorithm achieves better performance in predicting national GDP growth.By applying machine learning models,we can predict stability categories of different countries with a prediction accuracy of 90%.2)We perform a correlation visualization analysis of data-based environmental change and regional economic growth.We find that some developing countries have strong economic stability associated with water resources and carbon dioxide emissions,whereas the economic stability of developed countries is associated with per capita arable land.3)We propose new indicators.Compared with the current mainstream indicators,the new indicators are more focused on quantification of raw data,reducing the impact of conceptual abstraction indicators on forecast performance,which makes them more responsive to assessing a country's stabilization.Through actual forecasting effect analysis,the new ranking compensates for the prediction distortion defects caused by the abstract indicators in the world mainstream ranking when measuring regional stability and can satisfy the basic regional stability prediction function.Moreover,it may help us to better understand the prediction results with factor explanation.

Abstract:To overcome the instability of one single clustering result,we propose a new clustering ensemble method based on Dempster-Shafer theory (also known as belief function theory).In general,ensemble methods consist of two principal steps:generating base partitions and combining them into a single one;our method mainly focuses on the second step.After obtaining the base partitions in the first step,we convert them into an intermediate interpretation,which can be called a relational representation.We believe that the evidence source from the relational representations may be doubtful,which can be fixed by using the discounting process in belief function theory.After discounting the relational representations,we can combine them in the evidential level by different combination rules.Then,we can obtain the belief matrix or plausibility matrix from the fused relational representation,which can be seen as a co-association matrix between objects.To make full use of the transitive property between objects,we treat this co-association matrix as a fuzzy relation and make it the transitive closure to yield a fuzzy equivalence relation.The final partition is obtained by applying some clustering algorithms to the new co-association matrix.The experimental results show the stability and efficiency of our method.

Abstract:In order to improve the accuracy of face recognition and better display facial features,we propose the combination of LRC and CRC deviation based on mirror image for face recognition.First,the proposed method generates mirror images,then by merging the original face image and the mirror image to form a new mixed training sample.Finally,face recognition is performed by using LRC and CRC deviation.The new method increases the number of training samples,and overcomes the problem of variability from pose and illumination of the original face images.The experimental results show that the LRC and CRC deviation combined method significantly improves the accuracy of face recognition.

Abstract:The accuracy of geolocation on meteorological satellite data is critical for its application.The geolocation error sources of polar-orbit satellite are introduced,and then the research status on geolocation error estimation and correction is reviewed.Three mathematical methods of geolocation error correction,including Coastline Crossing Method (CCM),Node Differential Method (NDM) and Land Sea Fraction Method (LFM) are compared and discussed.Future researches in geolocation error correction are prospected.

Abstract:Lightning current amplitude cumulative probability is a necessary parameter to characterize the frequency of thunder activity as well as calculate the lightning strike flashover.Using data from lightning location monitoring system in Shaanxi province,this paper analyses the lightning current data during 2009-2012 in Yan'an area,and compares various calculation formulas of cumulative probability for lightning current amplitude.Then the density formula is introduced,and an example is given to show the difference between the calculated and measured cumulative probability of lightning current amplitude.The least square curve fitting method of CFTOOL in MATLAB is employed to find the best fitting parameters,therefore a more precise calculation formula for lightning current amplitude cumulative probability is established.The lightning current data in 2013 is used to verify the proposed formula.Results show that the average current amplitude of positive lightnings is higher in value and less concentrated in distribution than that of negative lightnings.The distribution curve of lightning current amplitude cumulative probability is relatively flat for positive lightnings and steep for negative ones.As for the simulation results by various calculation formulas,the regular formula has the biggest error,while the formulas recommended by IEEE Std and CIGRE have similar variation trends and shapes compared with measured curves.The fitting is best with parameter α being 36.04 and β being 4.349;the fitting error is ranged in -0.025 to 0.018 when the lightning current (I_c) is from 0 to 150 kA,and decreased to close to 0 when I_c is more than 150 kA.Thus a more precise formula is established to calculate the cumulative probability of lightning current in Yan'an area,which is then verified through test.