Research On Personalized Sequence Recommendation Algorithm Based On Knowledge Enhancemen

The sequential recommendation system considers the sequence of user interactions in the sequence modeling stage and can capture the dynamic interest preference of users better than the traditional recommendation algorithm.A key part of the sequential recommendation system is to recommend new items to users by learning from the collaborative signals in the history of user and item interactions.However,when the recommendation system is faced with the problem of sparse data,users’ interests and preferences cannot be accurately captured only by learning users’ historical interaction information.At the same time,different users have different interest patterns.How to better capture users’ personalized interest preferences in the sequence modeling stage is also an urgent problem to be solved.Knowledge graph is a large-scale structured knowledge system represented by triples formed by the structure of human knowledge.Introducing it into the recommendation system in the way of providing additional information can alleviate the problem that the existing sequential recommendation system cannot achieve the expected recommendation effect under the scenario of sparse data.However,the existing knowledge graph fusion method will introduce a large amount of information that is not highly relevant to the recommendation system in the project embedding stage of the recommendation system.The accuracy and operation efficiency of the recommendation system will be affected if all the information is introduced into the recommendation system.To this end,this thesis proposes a relational perception sampling method.Before the knowledge graph is introduced into the system,the relationships and entities in it are scored once for their relevance to the system,and only those interaction relationships with high relevance to the recommendation system are retained.At the same time,this thesis designs an end-to-end training strategy,which combines the process of relational perception sampling with the subsequent training process of recommendation system,so as to dynamically adjust the sampling process of knowledge graph according to the feedback of recommendation results to achieve the best results.In this thesis,extensive experiments are carried out on open data sets to prove that the knowledge graph relational perceptual sampling method proposed in this thesis effectively improves the performance of the sequential recommendation system combined with the knowledge graph.Meanwhile,the influence of the number of items retained by the knowledge graph in the sampling method on the results of the recommendation system is studied through comparative experiments.The research shows that users’ interest patterns can be roughly divided into two kinds of interest patterns: local dynamic preference and global stable preference.Different users have different preferences to the two interest patterns.To solve the problem that the existing sequential recommendation system cannot accurately capture the user’s personalized interest pattern,this thesis proposes a knowledge-enhanced personalized sequential recommendation algorithm.The convolution kernel in the convolutional neural network has translation invariance,which can capture the local dynamic interest features of users well by means of sliding window.Self-attention mechanism is successfully applied to capture the global interest features of users because it can set different focus points in different scenarios and avoid the problem of gradient disappearance in the modeling of long sequences of recursive neural networks.In order to better model the personalized interest features of users,we designed a hybrid modeling module.First,convolutional neural network and self-attention mechanism were used to compute the two interest features of users in parallel,and then an adaptive dynamic fusion method was designed to integrate the two interest features dynamically according to the different personalized interest preferences of users.On the public data sets Movie Lens-1M and Amazon-book,the knowledge-enhanced personalized sequence recommendation algorithm proposed in this thesis is compared with the most advanced recommendation algorithms at present.Experiments show that the proposed algorithm has certain improvement compared with the baseline method in multiple evaluation indexes.This indicates that the knowledge-enhanced personalized sequential recommendation algorithm proposed in this thesis effectively improves the accuracy of the recommendation algorithm,and the ablation experiment verifies that the recommendation results of the proposed algorithm are interpretable to a certain extent.