KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos

Source

Core Claim

KuaiRand provides sequential recommendation logs with randomly exposed videos, making action exposure less confounded than ordinary recommender logs.

Action-Time-Series Notes

  • The time-series unit is a user interaction sequence over recommended/exposed videos and feedback signals.
  • The action is the item/video exposure; feedback signals are response observations or rewards.
  • It is much more suitable for action-response user models than one-step ad-click logs.

Foundation TSFM Relevance

Agenda slotVerdictEvidenceMissing pieces
Causal structure, counterfactuals, and controlpartially closesRandomly exposed videos give large-scale intervened interaction logs with item exposures as actions and 12 user feedback signals as outcomes.Recommendation actions are randomized exposures, not system-control interventions with long-horizon counterfactual rollout.
Time representation and irregular event streamsadjacentProvides timestamped sequential user interactions and behavior histories.Needs explicit irregular-time modeling contracts and event-duration semantics.
Benchmarks: what level of modeling is tested?adjacentSupports unbiased offline evaluation, debiasing, interactive recommendation, long sequential behavior modeling, and multi-task learning.Does not evaluate general latent-state TSFM forecasting or control utility.