OneTrans: Unified Feature Interaction and Sequence Modeling with One Transformer in Industrial Recommender

In recommendation systems, scaling up feature-interaction modules (e.g., Wukong, RankMixer) or user-behavior sequence modules (e.g., LONGER) has achieved notable success. However, these efforts typically proceed on separate tracks, which not only hinders bidirectional information exchange but also prevents unified optimization and scaling. In this paper, we propose OneTrans, a unified Transformer backbone that simultaneously performs user-behavior sequence modeling and feature interaction. OneTrans employs a unified tokenizer to convert both sequential and non-sequential attributes into a single token sequence. The stacked OneTrans blocks share parameters across similar sequential tokens while assigning token-specific parameters to non-sequential tokens. Through causal attention and cross-request KV caching, OneTrans enables precomputation and caching of intermediate representations, significantly reducing computational costs during both training and inference. Experimental results on industrial-scale datasets demonstrate that OneTrans scales efficiently with increasing parameters, consistently outperforms strong baselines, and yields a 5.68% lift in per-user GMV in online A/B tests.

💡 Research Summary

OneTrans addresses a fundamental limitation of modern industrial recommender systems: the separation of user‑behavior sequence modeling and static feature interaction into two distinct modules. This split hampers bidirectional information flow, fragments execution, and prevents unified scaling. The authors propose a single Transformer‑based backbone that jointly processes both types of data.

Unified Tokenizer
All raw inputs are first embedded and then divided into sequential (S) and non‑sequential (NS) subsets. Non‑sequential features (categorical and numeric) are either grouped and passed through group‑wise MLPs or projected once by a single MLP and split (auto‑split). This yields L_NS NS‑tokens, each of dimension d. Sequential features consist of multiple behavior streams (e.g., purchase, add‑to‑cart, click). Each stream is projected by a dedicated MLP to the same dimension d, then merged either timestamp‑aware (interleaved by time) or impact‑ordered (high‑intent actions first) with learnable