In order to improve the accuracy and interpretability of social recommendation algorithms in the process of modeling user historical behavior sequences in a session,and aiming at the problem that the existing social recommendation algorithms do not fully consider the introduction and learning of causal information,a social recommendation algorithm based on counterfactual reasoning at the session level was proposed.In order to introduce causal information,the counterfactual data generator was used to generate such data for a historical behavior sequence of the user's friends in the session,and the causal information of social relations was captured through the counterfactual data.In order to learn the causal information,a multi-task prediction module consisting of multiple loss functions was proposed on the basis of the factual and counterfactual data,and the causal information was learned in the process of training the model.Compared with the existing excellent methods on two real data sets,Douban and Deli-cious,the experiment proved that the proposed algorithm had been improved compared with the baseline method on the evaluation indicators Recall@K and MRR@K.
social recommendationcounterfactual reasoningcausal informationsession
维普
万方数据
