Research On Serialization Recommendation Technology Based On Knowledge Graph

With the rapid development of information technology,massive amounts of information are presented to us through the Internet,which makes the problem of information overload more and more serious.As an effective way for users to efficiently obtain interesting information under "information overload",recommendation algorithms have become a hot research topic.The core of the recommendation algorithm is to build a bridge between user’s interest and massive Internet information.The specific process can be abstracted as a two-tower architecture composed of the user tower and the item tower: it obtains the interest representation on the user tower side according to the user’s historical interaction information,obtains the item representation on the item tower side according to the item auxiliary information,and finally predicts the user’s interest in the item according to the correlation between the two to determine whether to recommend the item.At present,the recommendation algorithm based on the two-tower architecture has achieved great success,but there are still three problems that need to be solved urgently:(1)For the "user-item" historical interaction information on the user tower side,it is difficult for existing models to model long user histories and time-series features in historical sequences accurately.(2)For the auxiliary information on the item tower side,the existing models tend to ignore the correlation between auxiliary information,and it is difficult to avoid the information loss when multiple auxiliary information is fused into item representation.(3)The two-tower architecture usually crosses the information of the two towers to further improve the model effect.For the information crossing,the existing model is difficult to mine the correlation of the two towers.At the same time,too many cross parameters can easily lead to negative transfer between information.In order to solve the above three problems,the main contents of this thesis are as follows:(1)In response to problem 1,this thesis proposes a serialization recommendation model based on the time attention mechanism.The time series features are modeled by designing a time encoding function,and a combination method is designed to introduce the time series features into the attention mechanism to accurately model complex dependencies in long sequences from a time perspective.(2)In response to problem 2,this thesis proposes a knowledge graph representation learning model based on multi-granularity fusion of heterogeneous information,mines the correlation between auxiliary information through knowledge graph representation,and designs a multi-granularity aggregation module based on graph neural network and field-aware factorization machine to combine item auxiliary information from different granularities to avoid information loss.(3)In response to problem 3,this thesis proposes a multi-task learning framework to model the above-mentioned serialized recommendation model and knowledge graph model,and builds a serialized recommendation model based on knowledge graph.It first mines the correlation between two towers(that is,between two model parameters)and avoids negative transfer between information by designing a non-invasive parameter sharing mechanism combined with an attention mechanism,and then uses an alternate learning method to complete information interaction.Finally,through experiments,this thesis verifies that the serialized recommendation model based on knowledge graph can effectively improve the recommendation effect.(4)This thesis designs and implements a personalized recommendation system based on the serialized recommendation model based on knowledge graph,and conducts recommendation tests in movies and e-commerce scenarios to verify the effectiveness.