Research On The Method Of Movie Recommendation Based On Asistant Neighbors Sampling And Long Short Term Interest

In recent years,with the rapid development of the film industry and the arrival of the streaming media era,a large number of movies available for online viewing are produced worldwide every year.However,too many movies makes it very difficult for users to choose.The arrival of recommendation system is precisely to get this problem fixed.It uses recommendation methods to filter out movie that is of interest to users from a large number of movies and improve user experience.Because of the rich semantic relations between entities and entities contained in the knowledge graph,it has been applied in movie recommendation system as assistant information,which has improved the performance of the recommendation.Among them,knowledge graph recommendation method based on propagation learning is an important area of knowledge graph recommendation and it has been widely used.Among the existing knowledge graph recommendation methods based on propagation,random sampling is used when aggregating neighbor nodes.Its randomness leads to the introduction of many unrelated nodes,which brings a lot of noise.In addition,the method of knowledge graph propagating recommendation in existence is not accurate enough for users,and does not take the short-term and long-term interests of users into consideration,and cannot capture users’ recent concerns and long-term interests.To address the issue of inaccurate representation of movie items caused by the randomness of neighbor sampling in existing methods,this thesis proposed a Movie Neighbor Aggregation Method based on Asistant Neighbor Sampling.The results of the two parts were input into the inner product function for calculation to obtain the probability of users being interested in movies.Movies with high probability would be recommended to user.The two parts are now introduced as follows:1.To solve the problem of inaccurate item representation caused by random neighbor sampling in existing methods,this thesis proposed an item neighbor aggregation method based on assistant neighbor sampling to solve the above shortcomings.First of all,this thesis used Internet Movie Database(IMDb)to assist neighbor sampling,and used the ranking information on IMDb as a sampling reference to purposely screen out the more important neighbors around the movie entity.Then,the filtered neighbor vectors were represented using the Trans R method,and aggregated to the vector of the central node to generate a single-level item representation.Finally,the single-level item representation was extended to the multilevel item representation,and the neighbor nodes around the nodes in this layer were also aggregated to the nodes in this layer to achieve the accurate representation of movie item nodes.2.Aiming at the problem that the user representation in the existing recommendation methods based on propagation learning was not accurate enough,this thesis proposed a user representation method to mine users’ long-term and short-term interests.First,the neighbor sampling method sampled and aggregated the neighbor nodes of the movies that users had interacted with in the past.Then,for the movies that users had interacted with for a short time,input these items into the self-attention network to get their respective attention weights,and used these weights to weight and aggregate the items that users had interacted with recently.For movies that users had interacted with for a long time,these items were aggregated by mean aggregation.Finally,the user’s short-term interaction aggregation results and long-term interaction aggregation results were input into the aggregator for aggregation,and the accurate user representation of the user was obtained.This thesis verified the effectiveness of the method proposed in this thesis through Click-Through rate(CTR)prediction experiments and Top-K experiments on the public Movie Lens-20 M and Movie Lens-1M dataset.Benchmark experiments with existing benchmark methods showed that this method performs well in AUC,F1 values and Recall.And in the end,two sets of ablation experiments were conducted to verify the advantages of each part of the proposed method in this thesis.