Font Size: a A A

Research On Matrix Decomposition Context Aware Recommendation Method Based On Deep Learning

Posted on:2021-05-02Degree:MasterType:Thesis
Country:ChinaCandidate:Z LiuFull Text:PDF
GTID:2428330611953109Subject:Computer application technology
Abstract/Summary:PDF Full Text Request
In order to make the existing recommendation system meet the intelligent recommendation needs of users in highly sparse data environment,more and more scholars have studied and improved the context aware recommendation method.In the field of situational awareness recommendation,situational information includes two categories,which are static situational information and dynamic situational information.Static situation information is the inherent attribute of users and projects,and its value range can be pre observed by the recommendation system.Dynamic situational information is the comment or annotation information that contains user preferences in the interaction process between users and projects.The existing recommendation methods do not fully consider the impact of situation information on user rating,resulting in low accuracy of recommendation,unable to meet the needs of intelligent recommendation.Aiming at this kind of problem,this paper studies the method of situational awareness recommendation,and proposes two kinds of situational awareness recommendation models for two kinds of different situational information.Firstly,aiming at the static situation information,this paper introduces the optional static situation information as an independent characteristic factor into the traditional matrix decomposition model,and outputs the accumulated value after interaction with user factor and project factor as the prediction value,thus proposes a matrix decomposition recommendation model integrating the static situation information.In addition,for dynamic situation information,the traditional recommendation system is difficult to extract its dynamic hidden features effectively,and there are problems such as the separation of scoring information and situation information.In this paper,deep learning technology is introduced to learn the deep nonlinear features of scoring and dynamic review data,and through the deep fusion layer to cross fuse the multi-source feature vectors,a multi-source feature hybrid recommendation model based on deep learning is proposed,which effectively solves the problem of lowaccuracy caused by the separation of dynamic situation information and scoring information.The specific work of this paper is as follows:1? This paper proposes a matrix decomposition recommendation model which integrates static situation information.The situation information is introduced into the traditional matrix decomposition model as an independent dimension.Through the interaction and accumulation of situation conditions,users and project factors respectively,and considering the different sensitivity of different users and projects to the situation information,the global situation factor vector is added to balance the prediction deviation caused by the different sensitivity of users and projects.Finally,the final predicted score was obtained by training model.2?A multi-source feature hybrid recommendation model based on deep learning is proposed.Firstly,multi-layer perceptron network is used to learn the potential non-linear features of sparse scoring data.Then we use the Bert model and self attention mechanism to learn the potential features of the project review set.Through the long and short-term memory network,we can learn the dynamic preference characteristics of users in the user comment set.Different from the traditional method,which directly uses inner product for scoring and prediction,this model constructs a deep fusion layer,which makes full use of the correlation between scoring features and comment features through the cross features of different levels and different granularity,so as to further improve the accuracy of the model.
Keywords/Search Tags:Recommendation algorithm, matrix decomposition, deep learning, situational recommendation
PDF Full Text Request
Related items