User Preferences Based Recommender Systems Using Deep Learning

The development of the 21 st century is accompanied by the outbreak of the information revolution,and the amount of global data has grown exponentially.Excessive information makes people dizzying and difficult to find information they need.In order to alleviate the problem of information overload,the recommender system becomes an increasingly important means of personalized information filtering.Deep learning has made breakthroughs in many areas.This thesis applies it to recommender system and achieves significant results.At the same time,knowledge graph greatly improves the efficiency and quality of people's access to knowledge.As a knowledge base rich in effective information,the knowledge graph can also provide a lot of auxiliary information to the recommender system to improve the recommendation performance.In the recommender system,accurate modeling of the user's preferences is crucial to the quality of the recommendation.With the help of deep learning methods,this thesis mainly focuses on the modeling of user preferences and completes the following two research work:(1)In many online websites where a recommender system is applied,the interactions between the user and the system are usually organized into sessions by time period.Most of the existing methods assume that the sessions are independent of each other.They either ignore long-term information from the user's history session or treat user preferences in the same session as static.In fact,most online websites record the user's choices.There is a connection between the same user's sessions;and even within the same session,the user's preferences will change.The existing methods neither fully utilize the user's historical session information,nor represent the user's behavior in the real environment in a finegrained manner.Therefore,this thesis proposes a new framework-Recurrent Memory Networks(RMNet),which models the user's long-term preferences.The key components in the RMNet are the session-level preference drift unit and the attention retrieval unit,which help to explicitly represent long-term user preference shifts and short-term preference changes,respectively.This thesis not only demonstrates the advanced nature of RMNet relative to recent related research work from a theoretical view,but also carries out extensive experiments on a real job recommendation data set.The experimental results prove that RMNet has achieved substantial gains over the most advanced baseline methods.(2)In order to solve the sparsity and cold start of collaborative filtering,researchers often use side information(such as social network or item attributes)to improve recommendation performance.This thesis uses knowledge graph as a source of side information.In order to solve the limitations of existing recommender systems based on knowledge graph,we proposed Ripple Network,which is an end-to-end framework that naturally incorporates the knowledge graph into the recommender system.Similar to the ripples spreading across the surface of water,Ripple Network automatically iteratively expands the user's potential interest along the knowledge graph to simulate the spread of user preferences over the set of knowledge entities.The multiple“ripples”activated by a user's historically clicked items are superposed to form the preference distribution of the user with respect to a candidate item,which could be used for predicting the final clicking probability.We conduct extensive experiments in three scenarios for movies,books,and news recommendations.Ripple Network achieves better recommendation performance than the baselines.At the same time,Ripple Network provides a new perspective on the interpretability of recommendations.