用于捆绑推荐的双视图对比学习OA

Bundles can satisfy multiple user preferences at once.Most existing bundled recommendation models endeav-our to capture user preferences from different perspectives.However,these models encounter two problems:(1)The user preference for potential interaction bundles cannot be fully captured.(2)The correlations among bundles are not adequately extracted.To address these problems,the paper proposes a dual-view contrastive learning for bundle recommendation model(DCLBR).Specifically,DCLBR introduces an item-level hypergraph to capture the user preference for potential int-eraction bundles in item view,and an attention network is adopted to adaptively aggregate the representations of correlated items to obtain bundle representations.Then,this paper generates a bundle-level weighted graph to mine correlations among bundles in bundle view.In addition,in order to make the bundle more compatible with the preferences of users,the paper generates negative and positive bundles by performing data augmentation based on masking of important and unim-portant items,respectively.Contrastive learning is leveraged to make the final bundle representation adaptive to the impor-tance of the items.Extensive experiments on three public datasets show that this model outperforms baseline models.