融合对抗网络和对比学习的三重桥梁跨域推荐OA

A triple bridge cross domain recommendation model(ACTBCDR)that integrates adversarial networks and con-trastive learning is proposed to address the problems of negative preference modeling deficiency and single user prefer-ence transfer strategy in existing meta learning cross domain recommendation methods.Positive and negative feature encoders are designed to obtain users'positive and negative preferences,in order to capture more comprehensive user preferences.To solve the problem of a single user preference migration strategy,a triple bridge of personalized preferences,source domain common preferences,and target domain common preferences is constructed,and a dynamic gating mecha-nism is adopted to achieve multi-granularity feature fusion.A domain adversarial training module is designed that achieves cross domain user representation alignment through gradient reversal layer(GRL),and the common preference bridge is mapped between the source and target domains to the same space by comparing loss constraints,reducing domain differences.Experiments on the benchmark datasets of Book,Music,and Movie in Amazon have shown that ACTBCDR performs well in MAE,RMSE,AUC,and NDCG@10.The indicators exceed the baseline method by 7.28%,8.29%,1.68%,and 6.91%respectively,verifying the effectiveness of the model.