Research On Ranking Algorithm Of Recommender System Based On Neural Network

With the rapid development of information technology and mobile Internet,the problem of information overload has become more and more serious,which raises two prominent contradictions:the difficulty to extract useful information from massive information and the difficulty for information producers to accurately convey personalized information to the users who are interested in it.According to the user's interest,behavior and the scene,the recommender system pushes to the user the recommended content,ranked by relevance to user's interest,thus effectively solving the two contradictions caused by information overload.This thesis reviews the research progress of the recommender system and summarizes some problems existing in the current research.Collaborative filtering is one of the most classic recommender algorithms,however,it is particularly difficult to recall enough data in the recommender system having sparse data.Logistic regression is another classic sorting algorithm,but it is difficult to cross sparse features.Factorization Machine model can generate binomial crossing features but can do nothing to get polynomial crossing features.Deep learning model can generate polynomial crossing features via embedding.However,the embedding method is inflexible,and the polynomial crossing features do not contain the original crossing term of eigenvalues.In this thesis,we propose a new recommender model based on a dynamic feature embedding and polynomial crossing features method of a diamond structure neural network.The model is distinguished from other frameworks in three characteristics.Firstly,it is an end-to-end model.It does not need to artificially construct crossing features and it can automatically generate crossing features.Secondly,the feature embedding module dynamically select the embedding dimension according to the number of the domain,and simultaneously performs different domains.Thirdly,a diamond structure neural network is used to generate polynomial crossing features,not only to preserve the original crossing term of eigenvalues,but also to make the polynomial crossing features abstract by iterative operation of "summation-diffusion-summation" turn.Last but not least,the experiments show that the proposed model has stronger predictive and ordering ability than the traditional recommender model as well as the common deep neural network.