Research On Personalized Recommendation Method Based On Knowledge Graph

In the era of big data,all platforms want to use the data and information to use the recommendation system to calculate and better display the results to obtain benefits.Therefore,how to dig deeper into user preferences has become a current research hotspot.Nowadays,more algorithm models emerge in an endless stream,and research shows that it is necessary to introduce auxiliary data into the recommendation system.In recent years,the knowledge graph has received more and more attention due to its comprehensive auxiliary data.It usually expresses the semantic information of the entity and the relationship between the entities in the form of triples,and can accurately express the semantic information of the interaction between the user and the item.By exploring the internal connections in the knowledge graph,the connectivity between users and items reflects their underlying relationships,and these relationship information gives the recommendation system the ability to reason and explain.Introducing multi-modal data feature fusion into the recommendation system can better mine user preferences.In response to the above problems,this article mainly studies how to make personalized recommendations to users.The main research contents are summarized as follows:(1)Designed a method to explore the spread of user preferences.This method mainly uses the triples in the knowledge graph to clearly show the multi-step interrelationship,which is one or more paths of user-item pairs.This paper explores these inference paths to achieve comprehensive inference and analysis of the recommended results.In the knowledge graph,the path from the user to the item is formally defined as a sequence of entities and relationships.Given the embedded item and the user's preference set,each triple in the preference set is assigned an association probability by comparing the item and the head entity and relationship of the triple.The association probability can be regarded as in the relational space The similarity between the measured item and the entity.After obtaining the associated probability,take the sum of the tails and weight with the associated probability to obtain the vector value,which is regarded as the first feedback,and finally the evaluation score is calculated by the formula.Get the user preference.(2)Design a method for recommending the interpretability of knowledge perception paths.The knowledge-aware path model includes three layers: embedding layer,LSTM layer and pooling layer.The user-item pair path of the user preference is used as the input of the embedding layer to obtain the embedded data set;then the embedded data set is input to the LSTM layer,the data is encoded,trained,and effective inference is performed in the layer model to obtain each Path prediction score;because the prediction score does not specify the importance of each path,this paper designs a pooling layer to join the pooling operation to aggregate the scores of all paths,use gradients to express the importance,and select the path with the highest score as An interpretable path is recommended.The top@K and pn@K results of the two datasets of movies and books are 15%,20% and 10%,20% higher than the best baseline model when K=4.The AUC results are higher than the best The baseline increased by 3.5% and 6.8%,respectively,and the ACC increased by 4.5% and 5.0%.(3)Designed a multi-modal feature interactive deep fusion recommendation method.This method first uses different methods to process and interact with multi-modal feature data,and convert the title data into sequence features;video and audio use PCA dimensionality reduction,input into two different DNN layers,and then combine the original feature data input Go to the deep fusion model and train the model.The deep fusion model is based on the x Deep FM model,which combines multiple methods for weighting,and finally predicts the final result.This method can not only perform high-level feature learning on implicit and explicit multimodal features,and transform feature interactions into trainable vector values,but also has the characteristics of breadth and memorable learning,and can perform multi-task training.After data set validation,the model's Score is 0.79363,which is0.02 higher than the best baseline model.