Graph collaborative filtering (GCF) is a popular technique for capturinghigh-order collaborative signals in recommendation systems. However, GCF'sbipartite adjacency matrix, which defines the neighbors being aggregated basedon user-item interactions, can be noisy for users/items with abundantinteractions and insufficient for users/items with scarce interactions.Additionally, the adjacency matrix ignores user-user and item-itemcorrelations, which can limit the scope of beneficial neighbors beingaggregated. In this work, we propose a new graph adjacency matrix that incorporatesuser-user and item-item correlations, as well as a properly designed user-iteminteraction matrix that balances the number of interactions across all users.To achieve this, we pre-train a graph-based recommendation method to obtainusers/items embeddings, and then enhance the user-item interaction matrix viatop-K sampling. We also augment the symmetric user-user and item-itemcorrelation components to the adjacency matrix. Our experiments demonstratethat the enhanced user-item interaction matrix with improved neighbors andlower density leads to significant benefits in graph-based recommendation.Moreover, we show that the inclusion of user-user and item-item correlationscan improve recommendations for users with both abundant and insufficientinteractions. The code is in \url{https://github.com/zfan20/GraphDA}.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)