Research On The Problem Of Classification Of Time Series Data

With the rapid development of information technology,the popular way of sending message has gradually changed from text and images to videos.Compared with text data and image data,video data has the characteristics of rich information content,large amount of data,complex structure and so on.In the coming Internet of things era,video data will become the basis of the interaction between people and objects or objects and objects.At the same time,the complicated underlying structure of video data may bring a new challenge to the task of data analysis and information mining.Video data is a series of image sequences,which have temporal dependence naturally.In the analysis of video data,a basic problem is object recognition and motion recognition,which can be regarded as a classification problem in the field of statistics and machine learning.In classical classification model,a large number of literature focused on the classification of cross-section data,which have a wide range of applications in practice.However,the video data is obviously different from the cross-sectional data.We must take the temporal dependence structure of the data into full consideration in order to get a robust classification model.In order to solve these problems,this paper focuses on the temporal data classification.In this paper,we use the Bayesian network to model the temporal dependence,which is a method based on conditional probability.In the classical classification model,the logistic regression is also a method based on probability,the existing research often combined these two.However,in the task of video data motion recognition,it is not enough to only use the classification method based on probability;even in more complex classification problems,such as multi-label classification,unbalanced data classification,we need a more general classification model.In order to solve this problem,this paper proposes a general framework that introduces Bayesian network into the general classification problem,which can deal with classification problems with temporal structure.This general framework is highly flexible,and can deal with many kinds of temporal dependencies and different classification cases.At the same time,the models in the framework are very interpretable.We introduce hidden Markov model for temporal structure,and evaluate the performance of classification by generalization error Err(f)=EL(Yt.Yt).A general framework based on general classifier and Hidden Markov Model is proposed.In this framework,we briefly discuss the extension of the model,which shows that the framework is highly flexible.We also propose two methods to solve the model in this framework,one is based on EM algorithm and the other is based on coordinate descent algorithm.Based on this framework,the single-label temporal classification,multi-label temporal classification and multi-label temporal classification with weighted loss are studied in this paper.Temporal data classification is not only used in video data,but also in financial market.In the study of single-label temporal classification,based on the analysis framework in Chapter 3,we extend the classical Support vector machine method to problems with temporal structure.The results of simulation studies show that the temporal data classification model is superior to the traditional classification model without temporal structure in prediction accuracy.In the prediction of the future direction of the stock market,the model can not only improve the prediction accuracy,but also can explain the market in a reasonable way.For more complex video data,we study multi-label temporal classification problem and multi-label temporal classification with weighted loss.Different From single-label classification,multi-label classification needs to consider the existence of different labels at the same time,which brings the issue of labellabel dependence.Under Hamming loss,we introduce the label-label dependence by considering the adjacent label variables as the input variables of classifier,and extend the single-label temporal classification to the case of multilabel.For multi-label classification,()pointed out that the methods above can not describe the correlation between labels well,they proposed a new weighted loss function,directly measuring the label-label dependence by a loss function.In this paper,the weighted loss function is introduced into the general framework proposed in Chapter 3,and the multi-label temporal classification problem is solved from another perspective.These two methods can be easily extended under the general framework of this paper.The simulation results show that the method considering both the temporal dependence and label dependence is better than the method ignoring at least one kind of dependence.In the video data processing,we use the activitynet data set.We use C3d deep neural network to preprocess the raw data to generate input variables,and use the natural language processing technology to process the language description of the video content to generate action tags and potential state variables.The experimental results of the video data motion recognition task also show that considering both the temporal dependence and label dependence can improve the classification performance much.In this paper,we exploit a new framework for classification with temporal dependence.To illustrate this framework in details,we apply it to singlelabel classification,multi-label classification and multi-label classification with weighted loss,including models,algorithms,theoretical properties and performances in both simulation and real data analysis.This framework is a novel methodology,trying to combine the advantages of deep neutral networks and statistical modeling to build a powerful and explanatory model,especially for complected data structure,such as video.