A Matching Recommendation Mechanism Based on Deep Learning and Topic Model

Abstract

In recent years, text recommendation has been widely used in various APPs as a key technology for users to quickly and accurately obtain relevant information. Traditional text recommendation cannot obtain the internal relationship between users and articles, and ignores the information generated by users. Therefore, this paper proposes a matching recommendation mechanism based on articles and comments. First introduce the word2vec word vector model, use the vector to measure the relative meaning between words, and construct the document vector and user distribution vector based on the word vector. Then, under the framework of the topic model, a joint deep learning method—long and short-term memory network LSTM, makes full use of the new model before and after the sentence to learn the document to update the word vector expression of the sentence and document vector. Among them, the conditional random field (CRF) model is added to train the tags to solve the problem of insufficient attention to key words. Finally, in the matching mechanism, the similar relationship among the topic distributions, the constructed document vector and the user vector are used for training. Compared with the current popular topic model TopicRNN method, topic word vector model LF-LDA method, topic vector-based text representation method and four methods of LF-LDA combined with Word2vec text representation, the experimental results show that the matching recommendation classification is obtained Improved and very robust, training time is greatly shortened, the algorithm in this paper is effective.