Research On Recommendation Algorithm For Fusion Of Deep Hidden Semantic Features Of Text

In today's Internet age,the amount of information and data is increasing exponentially,which brings us a lot of convenience and troubles.That is to say,when faced with a huge amount of data,it becomes more and more difficult to obtain the data you want efficiently,which is the so-called "information overload" problem.By analyzing the historical behavior of users,the recommendation system actively recommends information that users may be interested in from a large amount of data,which has become one of the powerful means to alleviate the problem of "information overload".Although the traditional recommendation algorithm has achieved good results,there are still many problems,such as cold start,sparsity of data,inability to use other knowledge level data and so on.There are many research directions of recommendation algorithm,among which how to fuse other data into recommendation system to help solve the problem of data sparsity is one of them,and fusing text type data is one of its hot spots.The main research content of this paper is to study how to fuse text type data to further improve the performance of recommendation algorithm.Because of the disadvantage of using topic model to process text,it can't perceive text context.Therefore,this paper uses advanced deep learning model to capture text features,and then fuses the learned hidden semantic features of items into matrix decomposition model PMF.In this paper,new models HANMF(Hierarchical Attention Networks Matrix Factorization)and HANMF+ are proposed.The specific work is as follows:In this paper,firstly,we analyze the current recommendation algorithm model which fuses text features,and introduce a recommendation algorithm model ConvMF which fuses deep learning technology.The innovation of this model lies in using deep learning technology to perceive text context information,which improves the performance of the recommendation algorithm model.It is proved that strengthening the perception of text context features is helpful to improve the scoring prediction performance of the model.Therefore,Bi-GRU,a bi-directional gated cyclic neural network of RNN family,is used to process the text data(project description documents).Because ordinary RNN has the problem of gradient dispersion,GRU is used instead of RNN,and bidirectional GRU can capture the sequence information before and after the text.Because the ability of capturing text features can be further improved by fusing document structure information,this paper joins the hierarchical attention mechanism network HAN to pay attention to the feature of word-level and sentence-level structure information describing documents.Then,the hidden semantic features of the text processed above are properly integrated into the probability matrix decomposition model PMF from the perspective of probability,and its key hyperparameter are adjusted to improve the accuracy of scoring prediction.In order to improve the generalization ability of the model,this paper further proposes a pre-training word embedding model HANMF+.Finally,six models of PMF,CTR,CDL,ConvMF,HANMF,and HANMF+ were compared on three real data sets.Among them,the HANMF model increased by1.24%,0.72%,and 4.26% in the three data sets compared with the ConvMF model,which verifies the effectiveness of the model proposed in this paper.Because there are many influencing factors in the HANMF model,this paper also analyzes the important parameters and influencing factors.