Research And Application Of Recommendation Algorithm Based On Multi-Objective Ranking And Knowledge Graph Enhancement

With the rapid development of the Internet and the increasing scale of data,multiobjective recommendation,which models and predicts multiple objectives of interest simultaneously with the help of multi-task learning,is gradually becoming mainstream.Compared with traditional single-objective models,multi-objective ranking models can consider user needs and item attributes more comprehensively and provide better recommendation experience and business value without significantly increasing the number of model parameters.At the same time,with the gradual popularization of graph data and the application of graph technology,knowledge graph,as a kind of information source containing a large number of heterogeneous features,has also been applied to the construction and learning of recommendation models.It has been shown that the introduction of knowledge graph as side information into the recall model can improve the prediction accuracy of the original model to a certain extent and significantly alleviate the impact of data sparsity.In order to be able to incorporate the knowledge graph into the multi-objective ranking model,this paper mainly describes the technical principles related to multi-task learning and the recommendation model based on the knowledge graph,focuses on the recommendation algorithm based on the knowledge graph embedding,and summarizes the deficiencies of the existing models in the semantic feature extraction of the knowledge graph and other association information mining.Considering the characteristics of the multi-objective ranking scenario,Knowledge graph Enhanced Muti-Objective Ranking(KEMOR),a model based on knowledge graph enhancement is proposed and implemented.The model consists of three parts: the left multi-objective ranking module selects the PLE model as the sub-model,which aims to jointly learn multiple objectives using the shared representation.The right knowledge graph embedding module utilizes the ripple propagation and neighborhood aggregation mechanisms in both Ripple Net and KGCN models to perform higher-order feature extraction for user entities and item entities in the knowledge graph,respectively,and complete the embedding representation of entity nodes with the help of the auxiliary task of predicting interaction probabilities.The two low-level representations are connected through the cross & compression unit,and the user hierarchical interest representation and the item higher-order aggregation representation extracted from the high level of the knowledge graph side are concatenated into the shared representation of the multi-objective ranking side,which effectively improves the completeness of the shared representation and thus achieves the effect of knowledge graph enhancement.In this paper,we construct fusion knowledge graphs and conduct a series of experiments on two real-world datasets,Ali-CCP and Trade-Up,which are suitable for multi-objective ranking.The experimental results show that introducing the knowledge graph as side information into the shared representation of the multi-objective ranking model with the model proposed in this paper can effectively enhance the recommendation performance,with 1.7% and 2.2% improvement in the AUC values under the two objectives of CTR and CTCVR in the Ali-CCP test set,and 3.5% and 1.7% in Trade-Up.The experimental results simulating different sparsity scenarios show that the incorporation of information on the knowledge graph side can better alleviate the data sparsity problem.The experimental results for the negative transferring problem show that both the shared representation incorporated by cross-compressed units and the direct concatenated high-level representation in KEMOR can improve the negative transferring phenomenon and obtain positive gains for all targets.