Research On Emotion Classification Based On Shapelet Algorithms

The quality of emotions determines people’s living conditions.Positive emotions such as happiness,satisfaction and joy can promote happiness,while negative emotions such as anxiety,sadness and disgust can even cause people to suffer from mental diseases and threaten their lives.How to improve the accuracy of emotion classification and improve the universality of its application in daily life has attracted more and more researchers’ attention.Traditional emotion classification based on physiological signals is mainly carried out by feature extraction and training classifier.The disadvantages of traditional methods are low accuracy and poor interpretability.Shapelet is a sub-sequence of time series.It can appear in any position of time series and has discrimination ability to distinguish different kinds of time series.Since the waveform of physiological signals are time series,shapelet-based algorithms can be used to analyze it.This thesis focuses on shapelet-based algorithms and physiological signals-based emotion classification methods,and combines them.The main research contents of this thesis are as follows:(1)Fast shapelet discovery algorithm uses symbolic aggregation approximation(Symbolic Aggregate approximation,SAX)to reduce dimension and symbolize representation of time series,which is easy to lose trend information of time series and lead to the decline of classification accuracy.To solve this problem,a fast shapelet discovery algorithm(FS-TSAX)based on trend feature representation(Trend Symbolic Aggregate approximation,TSAX)is proposed.FS-TSAX combines the fast shapelet discovery algorithm with the idea of TSAX,which improves the accuracy of shapelet classification.Firstly,trend characteristic symbolization method is used to represent time series,and then random mask is used to select candidate shapelets and information gain is calculated.Finally,the best shapelet is selected.Experimental results show that the accuracy of the proposed FS-TSAX algorithm is higher than that of the existing fast shapelet algorithm in multiple data sets,especially in some time series with obvious trend characteristics.(2)The traditional emotion classification based on physiological signals is not explainable,and the collect of EEG signal is inconvenient.To solve these problems,a pervasive multiphysiological signal-based emotion classification with shapelet transformation and decision fusion(PMSEC)is proposed.Compared with EEG-based emotion classification method,PMSEC adopts ECG(Electrocardiograph),GSR(Galvanic Skin Response)and RA(Respiratory Amplitude),which has strong universality and broad application prospect.In this method,the concept of shapelet transformation is introduced,and shapelet transformation and feature extraction are performed on ECG,GSR and RA respectively.Then,six sub-classifiers are constructed for different physiological signals and the decision level fusion is realized.Finally,the classification results of the experiment are analyzed from multiple perspectives.