Research And Application Of Recommendation Algorithm Based On Knowledge Graph

Recommendation system can effectively alleviate information overload and is widely used in e-commerce,news portals,social networking sites,digital libraries,and other fields.By analyzing the user’s historical interaction behavior,it helps users filter out uninterested information and match them with potential demand information to achieve personalized recommendation.Traditional recommendation algorithms,e.g.,collaborative filtering,learn users’ preferences by measuring the similarity of historical interactions.But as the number of items increases exponentially,it cannot alleviate the cold start problem.Due to the combination of KG recommendation algorithm can effectively use the topology of nodes for reasoning and recommendation,which can alleviate the problem of data sparsity.However,existing KG-based recommendation algorithms suffer from two problems.On the one hand,there are problems of inaccurate knowledge representation in the KG embedding and a single method in information aggregation.On the other hand,most studies focus on the sampling of users’ favorite items(positive samples)and have not fully explored the negative sampling strategy.Therefore,the research of this work are as follows:(1)For the problems of inaccurate knowledge representation in KG embedding and single method in information aggregation,we combine the recommendation algorithm and attention mechanism,i.e.,we propose a bias-based graph neural network recommendation algorithm(BGANR).Firstly,we consider that the weight of neighbor nodes is different when the entity node propagates and there is a bias between predicted value and real value.We introduce a bias-based attention calculation method to balance noise effects.Secondly,when the node and neighbor information are aggregated,we design a multi-channel fusion function to capture the high-order connectivity between nodes.Finally,we verified the effectiveness of the BGANR algorithm according to different evaluation indexes on three real datasets and also analyzed the interpretability of the relationship between items.(2)To address the lack of exploration of negative sampling strategy in KG,we propose a novel algorithm,dynamic negative sampling for recommendation with feature matching(DNSR),which consists of sampler and recommender.Firstly,the sampler obtains the embedding representation of the node through a simplified convolution network,we use multiple feature matching techniques to filter the representative nodes in the negative sample set.Then,to optimize the pairwise ranking relationship between positive and negative samples,we designed a recommender with reinforcement learning.This recommender can feedback the corresponding reward and punishment signals and encourage sampler to generate informational negative samples.Lastly,we choose three datasets for experiments,the results show that DNSR algorithm is better than the current classical algorithm.In addition,we also verified the performance of feature matching technology through ablation experiments.We designed and implemented a prototype movie recommender system based on Django framework for the above BGANR algorithm and DNSR algorithm.There are 24 figures,13 tables and 73 references.