UniDot: A Unified Network for Sequence Modeling and Feature Interaction in Large-scale Recommendation
2026-08-17 • Information Retrieval
Information RetrievalArtificial Intelligence
AI summaryⓘ
The authors developed UniDot, a new recommendation model that combines two common approaches: one that looks at features of users and items, and another that looks at users’ behavior over time. They do this by treating all input data as tokens and using dot products to measure interactions, which helps the model understand both feature relationships and sequences in a unified way. UniDot’s design shares information efficiently and speeds up predictions. When tested at a large industrial competition, their model performed very well, finishing second.
factorization machinesfeature interactionsequential modelsembeddingattention mechanismdot producttokenizationMLP-Mixercollaborative filteringconversion prediction
Authors
Rongcheng Lin, Yan Sun, Jamey Zhang, Guanglei Xiong, Ivan Ji, Xianjie Chen, Shujian Bu
Abstract
Industrial recommenders rely on two model families that have evolved largely independently: feature-interaction models over multi-field user/item features, and sequential models over user-behavior histories. Production systems couple them only loosely. To unify the two, we present UniDot, a novel architecture for post-click conversion prediction built from the factorization-machine (FM) point of view: the embedding inner product---which powers collaborative filtering and lets a recommender generalize to unseen user--item pairs---is the same primitive as attention's query dot key scoring, so a single dot-product of tokens can underlie both feature interaction and sequence modeling. UniDot tokenizes non-sequential fields and multi-domain behavioral sequences into one shared token space and stacks a single macro-block in which a token-mixing bus and a sequence-retrieval bus (item tokens cross-attending the histories) run in parallel and exchange state each layer through an MLP-Mixer fusion, while an FM Highway carries explicit per-layer dot-product interactions around the residual stack directly to the classifier. The sequence side is embedded once per forward pass and shared by all consumers, bounding inference latency. Trained with a dual sparse/dense (Adagrad + Muon) optimizer, an auxiliary conversion-delay head, and multi-path mutual learning, UniDot finished as the runner-up on the Industrial track of the TAAC KDD Cup 2026.