search

xiangwang1223 / neural_graph_collaborative_filtering

Neural Graph Collaborative Filtering, SIGIR2019
MIT License
781 stars 261 forks source link

Have you remove the current user-item edge in gcn propagation process? #22

Open yuanyuansiyuan opened 4 years ago

yuanyuansiyuan commented 4 years ago

Hello i want to know that have you remove the current user-item interaction in the gcn propagation process? Because when you test, the user-item interaction is not included in the graph, which will cause the inconsistent betwwen train and test data setting.

xiangwang1223 commented 4 years ago

We construct the bipartite graph based on historical interactions in the training dataset. The interactions in the testing dataset are "unseen" and should be excluded from the graph, otherwise, it will result in information leakage.

Here the historical interactions appearing in the graph only focus on information propagation. Hence, I kind of agree with your point of ''inconsistence training", and we are working on this issue.

Thanks for your insightful comments.