Research On Latent Factor And Recommender System

Considering the huge amount of data,recommender system has become the most effective way to solve the problem of information overload.The core of the recommender system is to help users find the content they are most interested in.With the growth of data amount,recommender algorithm plays a more important role in daily life.Personalized recommender algorithm is the heart of recommender system,so in recent years,the research on recommender algorithm is of great importance.In this thesis,we improved the neural collaborative filtering model for typical useritem rating prediction problem in recommender systems.Our new model reconstructs matrix factorization using deep learning technology,and represents hidden factors of user and item by combining embedding and neural network.At the same time,considering side information,auxiliary vectors are further introduced to help model user and item.During experiments,the latent-factor-based neural matrix factorization model we improved in this study presents a score around 0.804 measured in RMSE.Compared with the art-of-state non-negative matrix factorization model,the prediction accuracy of our model is promoted by 7%.Compared with the current art-of-state deep learning model,the prediction accuracy is improved by 4%,which is a significant improvement in useritem rating prediction field.In the research of model's robustness,we tried the idea of noise injection.Based on the neural matrix factorization model,we further propose a new version with good robustness.A noise layer is introduced into the model to simulate random perturbation,so as to increase the resistance of the model to the malicious attack in the actual environment.And the experiment results show that the new model with noise layer shows better robustness.Under the same condition,the success rate of the new model for malicious attacks is reduced by 22.5%.To solve the problem of model generalization,we made detailed theoretical analyses and experimental comparisons regarding model complexity and regularization methods.In the experiment,we found that appropriate model complexity and regularization methods can significantly improve the generalization of model.Finally,in order to accelerate the process of industrialization of our model,we also made an exploration regarding cold start and model training efficiency problems that may occurred when the model is put into production.During research,we used the idea of feature transformation to improve the content-based cold start solution.In the experiment,compared with the wide used most-popular-list recommender solution,the algorithm we proposed improves the hit ratio by more than twice.