SUREL+: Moving from Walks to Sets for Scalable Subgraph-based Graph Representation Learning

criteo-click contains a sample of 30 days of Criteo live traffic data, each corresponding to one impression (a banner) displayed to a user and whether it is clicked [1]. Each record has 9 contextual features that are aggregated into a 270-dimensional edge feature. There are 675 unique campaign banners and 6.1M users, consisting of a bipartite graph of 16.5M edges: 97% is used for training, and the rest is evenly split for validation and testing based on temporal orders. The task is to predict which campaign the user is most likely to click among 651 candidates. twitter-2010 is an industry-level social network with 1.5B user following relations [2]. An edge (𝑖, 𝑗) of this network indicates that user 𝑖 is followed by user 𝑗. 1% of Twitter users who follow 10 to 1000 accounts are randomly sampled for evaluation. The task is to recommend which account they will most likely follow among 1001 candidates.