Research On Recommendation Algorithm With Deep Reinforcement Learning And Knowledge Graph

In the era of "Internet +," with the continuous expansion of information,users' interests have become diverse,so how to make people find the information that meets their needs quickly and accurately in the massive amount of information has become an urgent problem to be solved at present.The recommender system has been used in the industry and academia as an effective means of information filtering,so the research of recommender system has become the most popular topic.Most recommendation methods generally adopt the combination of deep learning and collaborative filtering,which improves the recommendation performance to a certain extent.However,these methods still have the following four problems:(1)they cannot capture users' dynamically changing interests;(2)less consideration is given to the influence of users' micro-behaviors on recommendation performance;(3)lack of modeling of long-distance dependencies;(4)the connection between items is not considered when constructing the candidate item set.To alleviate the above problems,this dissertation proposes two recommendation models based on deep reinforcement learning and knowledge graph.The main work is as follows:1.A recommendation model based on deep reinforcement learning Actor-Critic is proposed.The user's historical interaction records are grouped in chronological order and entered into the model in the form of sessions.The input of the model includes two parts,namely the item sequence and the corresponding behavior sequence,and modeled the transformation mode of the two sequences separately.The gated recurrent unit used in the model has been improved to capture the importance of the current behavior.In addition,the self-attention mechanism can be used to model the dependency of long sequences considering the context and avoid the loss of information caused by the long session.The environment simulator is built to simulate the online environment,and the feedback information from users is obtained,so as to adjust the recommendation strategy.The model is trained by the method of deep deterministic policy gradient.2.A recommendation model based on deep reinforcement learning and knowledge graph is proposed.Based on the above model,this model uses knowledge graph to construct candidate set related to user preference,which reduces the cost of calculation and improves the accuracy of recommendation.Knowledge graph contains the relationships between entities,captures the higher-order connectivity of objects,it selects items related to user preferences,and filters out irrelevant items.Experiments conducted on two datasets,JData and KKBox.The experimental results prove that adding user behavior information can fine granularity mining users' interest,self-attention mechanism can capture long-distance dependency,and constructing candidate item set using knowledge graph as auxiliary information can improve recommendation performance.