Research On Representation Learning Based Multi-view Dimensionality Reduction Method

With the rapid popularization of social media centered on user content production,Douyin,Kuaishou and Weibo have become the most commonly used software in people’s daily lives.Most of the massive data generated by these social media in real time are multimedia data such as short videos and images.The high-resolution and high-dimensionality brings great difficulties to traditional machine learning and data mining methods.In addition,most of the data on social media is unlabeled,except for very few data with user-personalized labels.Data labeling of these massive high-dimensional data will consume a lot of manpower and material resources.Therefore,how to reduce the dimensionality of these massive unlabeled high-dimensional data has become an urgent problem to be solved.With the rapid development of information technology,especially the enhancement of data feature extraction and collection methods,the above social media data can often be described from multiple perspectives,thus forming multi-view data.For example,a piece of text can be described in the languages of different countries;a picture can extract features from different angles such as shape,texture,and color;different short videos report the same event.Multi-view data can usually describe the target more comprehensively,which is beneficial to overcome problems such as occlusion and illumination.At the same time,on the basis of high dimensionality,multi-view data presents the characteristics of large scale and multi-source heterogeneity,which further brings difficulties to traditional machine learning and data mining algorithms,resulting in the dilemma of "data that is difficult to use".Therefore,this thesis aims at dimensionality reduction of unsupervised multi-view data,and studies how to effectively fuse and reduce dimensionality of multi-view data at the same time.The main research contents are as follows:Firstly,an adaptive multi-view dimensionality reduction method based on graph embedding is proposed.This method adaptively learns a shared similarity matrix to maintain the correlation of data between different perspectives after dimensionality reduction,and uses the distance between high-dimensional data to constrain the similarity matrix,so that the multi-view data after dimensionality reduction maintains high dimensionality the structure of the data.In addition,the reconstruction error term is added to reduce the loss of information as much as possible in the data after dimensionality reduction.This thesis compares this model with other models,and the experimental results show that the reduced-dimensional representation can achieve better performance on clustering/recognition tasks.Secondly,a deep multi-view dimensionality reduction method based on autoencoder is proposed.The model uses an autoencoder to extract its feature representation for each view separately,and on this basis,a graph convolutional network is introduced to learn the relationship between samples.At the same time,the Hilbert-Schmidt criterion is introduced to ensure that the learned low-dimensional data from multiple perspectives are complementary,thereby improving the discriminativeness of low-dimensional representations.Experiments verify that the reduced-dimensional representation can achieve better performance on clustering/recognition tasks.