Tencent UNI-REC Challenge Solution Uses Field-Aware RankMixer

This research paper describes the solution developed for the KDD Cup 2026 Tencent UNIREC Challenge, which focused on predicting the probability of conversion (pCVR) for target advertisements. The challenge required a model capable of jointly processing complex multi-domain user behavior sequences and various non-sequential multi-field features. The proposed model, named Field-Aware RankMixer (FA-RankMixer), incorporates a dual-stream bilinear fusion architecture. It begins by employing target-aware DIN (Deep Interest Network) modules to extract user interests from different behavior domains, distinguishing between recent and earlier interests for longer sequences. Semantic tokens are then formed based on feature fields and behavior domains, which are processed by RankMixer blocks for cross-token interaction. The architecture further includes a shallow MLP stream that complements the deeper RankMixer stream, with a group-wise bilinear module fusing their representations. This comprehensive approach secured the ninth position on the official leaderboard for the challenge.