Add bipartite graph support to `structured_negative_sampling` · pyg-team/pytorch_geometric#7799

(6 留言) (1 反應) (1 負責人)Python (19,985 star) (3,514 fork)batch import

featurehelp wantedsamplerutils

描述

🚀 The feature, motivation and pitch

Hello, First and foremost, I apologize for potentially misplacing this issue in the wrong section. When working with the negative sampling part and the necessity to use the function "structured_negative_sampling" there might be inaccuracies when dealing with a bipartite graph. This issue becomes more evident if the number of nodes on the right side of the bipartite graph exceeds the number of nodes on the left side. For instance, consider the movie dataset, which consists of 6000 users and 4000 items, often employed in link prediction and recommender systems. I'm uncertain whether adding this feature falls within the scope of this library. However, I bring it up because, in the contribution section, you mentioned that we should address any uncertainties with you.

Alternatives

No response

Additional context

No response

貢獻者指南

技術棧: python
領域: machine learningdata
議題類型: feature
難度: 3
預計時間: half day
活動狀態: active
清晰度: clear
前置要求: Basic understanding of PyTorch GeometricKnowledge of negative samplingFamiliarity with bipartite graphs
新手友善度: 45
研究方向: Examine the current implementation of `structured negative sampling` in `torch geometric.utils.negative sampling`. Identify how it handles bipartite graphs and where the inaccuracies arise. Suggest modifications to support bipartite edge types, possibly by allowing separate source and target node sets. Discuss with maintainers on the issue #7799 to align on scope and implementation strategy.