Research On Recommendation Algorithm Based On Knowledge Graph

With the acceleration of technological revolution and industrial change,information is exploding.The birth of recommendation system has provided a powerful tool to solve the problem of information explosion.It can push information that may be of interest to users just by analysing their historical behaviors,which helps users save time in finding information to some extent.In the recommendation system,recommendation algorithms such as factorization machines cross-combine different features and treat each user-item interaction information in isolation,i.e,they cannot explicitly exploit the implicit associations between users and users and between items and items.This results in obtaining feature embeddings that are not optimal and cannot overcome the item cold start problem.Therefore,this thesis proposes multimodal knowledge graph embedding based neural factorization machines(MKGNFM),the model consists of a multimodal knowledge graph embedding part and a recommendation part,and uses a joint training approach to predict the user-item interaction probability by feeding multimodal features into the downstream recommendation task.The multimodal knowledge graph embedding part aims to obtain rich feature embeddings and overcome the problem of isolated treatment of samples by feature intersection models.Firstly,for structured knowledge,this thesis adopts the knowledge graph embedding model Hole,which preserves the structural information in the knowledge graph while obtaining entity and relationship embeddings.Second,considering the diversity of data types in multimodal knowledge graphs,this thesis uses the SIF algorithm to calculate the weighted average of sentence word vectors in the corpus based on the word vectors trained by the Glove algorithm,and uses them as text knowledge embedding features.The recommended part adopts the neural factorizer model,which has the advantages of both the second-order cross-linearity of the factorizer machine and the high-order crossnonlinearity of the deep neural network,and its bi-interaction layer can well integrate the two modal features generated by the embedding part of the multimodal knowledge graph.In order to verify the effectiveness of the model,this thesis conducts an empirical study of the recommendation algorithm by combining the Movie Lens-1M dataset and the movie synopsis information crawled on the IMDB website.The comparison test between MKGNFM model and the baseline model,the ablation test of the effect of different modal data on the recommendation results,and the hyperparameter test of the model recommendation module are conducted respectively.In this thesis,ROC-AUC and PR-AUC are used as evaluation indexes.The experimental results show that the MKGNFM model surpasses the general baseline model,the embedding of multimodal knowledge graph information improves the effect of the feature intersection model to a certain extent,the architecture of the model has certain rationality and feasibility,and the model is effective in improving the performance of the recommendation system.