Research And Implementation Of Recommendation Algorithm Based On Knowledge Graph

Recommendation system is an important growth engine of Internet companies.The recommendation system recommends items of interest to users by analyzing the historical behavior data of users.Excellent recommendation service can enhance the stickiness of users and improve the conversion rate of users.Recommendation algorithm is the core of recommendation system,which is responsible for data processing and the construction of recommendation model.Traditional recommendation algorithms have some problems,such as sparse data,cold start,and diversity of recommendations.Knowledge graph can effectively solve these problems because of its rich knowledge content and strong relationship processing ability.Recommendation algorithm based on knowledge graph is a research hotspot in recent years.Most of the existing work is based on graph neural network to build recommendation model,and the effect of recommendation is significantly improved compared with traditional algorithms.However,there are still some deficiencies in the existing recommendation models.Firstly,the existing recommendation models do not make good use of the timing information of the interaction between users and items in the graph,and do not model the user's interest over time.Secondly,the existing recommendation models do not use the relationship between items and attribute entities in the knowledge graph to model user's interest in a more fine-grained manner.Finally,in the dynamic recommendation scenario,when the knowledge graph needs to be updated frequently,the existing recommendation model has the problem of high training time and cost.This paper proposes a new recommendation algorithm based on knowledge graph and two recommendation model updating methods suitable for different scenarios to solve these problems.The main contributions of this paper are as follows:1)Based on the graph attention network,this paper proposes an attention recommendation algorithm integrating temporal and relational information-SRGAT.SRGAT uses the attention mechanism to process the timing information of the interaction between users and items,and uses the relationship between items and attribute entities to model users' interests.After the embedded vectors representing users and items are obtained,the layer aggregation of each layer of embedded vectors is carried out to obtain the feature vectors,and then the deep neural network is used to cross and combine the feature vectors to obtain the user's prediction score of items.2)This paper proposes two efficient recommendation model updating algorithms,namely ILU algorithm based on incremental learning and RDU algorithm based on reverse diffusion.ILU adopts the combination of regularization and data replay.It can train a new model only by using new data and some old data,so as to update the parameters of the model? RDU adopts the method of local updating,which only needs to update some model parameters without retraining the model.The performance of RDU update algorithm is not as good as ILU,but its update efficiency is higher than ILU.The two updating methods can be combined to update the model quickly and effectively.3)The proposed recommendation algorithm and recommendation model updating algorithms are compared on multiple public datasets.The experimental results verify that the recommendation algorithm proposed in this paper can improve the effect of recommendation,and the model update algorithms proposed in this paper could improve the update efficiency of the model on the premise of ensuring the performance of the updated model.