Attention Pattern Mining And Application Of Multivariate Student Behaviors Based On Time-Frequency Analysis

The study of the diverse behaviors of students on campus has been a topic of interest.With the development of information technology,campus behaviors of students become more complex and diverse.The thesis introduces the concept of "attention",based on campus behavior data,into the characterization of multiple campus behaviors of students,and studies the group division of students based on the attention characteristics of their multiple behaviors.The main research content includes the following three aspects:(1)The time-frequency feature extraction of students’ multiple behaviors was explored.First,an attention-based signal model of students’ multiple behaviors was proposed.The model abstracts multivariate behaviors as multi-channel signal sources,and considers individual attentional streams as non-smooth random signals.Then,the Hilbert-Huang Transform is introduced to express the time-frequency features of attention in students’ multivariate behaviors as a time-frequency spectrogram.(2)A model for the characterization of students’ multivariate behaviors based on attention is proposed.The model combines the stability,span,distribution and transferring qualities of attention and the time-frequency characteristics of multiple behaviors.On the one hand,it provides a unified characterization of multiple behaviors from the perspective of attention qualities.On the other hand,it can represent the patterns in students’ multiple behaviors from the perspective of signals.In the experiments,the model was used to extract the attention features in multiple behaviors and applied to K-Means,HAC and GMM clustering.The results show that the clustering effect using the attention model is better than those using traditional statistical features.It means that the representation model has good ability to distinguish multivariate behavioral patterns.(3)A deep clustering approach that incorporates attention quality and timefrequency features is proposed to support group division of students based on their multivariate behaviors.The approach uses the multidimensional time series of attention quality and its time-frequency spectrogram,and improves the macroscopic and microscopic differentiation of multivariate behaviors depending on a stacked encoder-decoder.Experiments show that the approach has advantages on the indices SC,CHI,and DBI over existing behavior clustering approaches.In addition,the correlation between the academic performance of students and their attention model is analyzed to further verify the reasonability of the proposed attention model for multivariate behaviors.