UniCon: A Unified Context-Centric Modeling Paradigm for CTR Prediction
Jiajun Cui, Zhengqi Xu, Fan Zhang, Zhangteng, Gu Tang, Honghong Zhu, Mengxi Wu, Yulin Liang, Xingxing Wang
cs.IR
2026-09-03
UniCon treats each exposure event's co-displayed items, intent, and environment as one context unit, runs history and candidates through the same hierarchical attention, and gains 0.0139 offline AUC plus 3.09% RPM in Meituan search ads.
Industrial CTR models increasingly unify sequential and non-sequential signals inside one backbone, but the input organization is inherited from feature-engineering practice: history is a sequence, candidates and request signals are non-sequential features scored pointwise. That split misaligns with how decisions happen. Each action occurs inside an exposure context; clicking the cheap nearby coffee shop says little about unconditional preference when the co-displayed alternatives were pricier or farther. Flattening everything into a token sequence conflates co-display within one impression with temporal adjacency across impressions.
UniCon changes the basic unit to the context unit: all items jointly displayed in one exposure event, plus intent, query, scene, and device signals. An input is a sequence of observed historical context units followed by one candidate-initialized target latent context unit.
Intra-context attention captures Locality, the competition and complementarity among co-displayed items; inter-context attention captures Dynamics, how interests and environments evolve. The two levels alternate in stacked blocks. Historical item tokens carry logged click feedback; candidate tokens fill that slot with a learnable placeholder, so the schema aligns without label leakage. Exposure and absolute-position auxiliary tasks pull the target unit toward the latent structure of the eventual display list.
The engineering side carries it to production: padding-free variable-length attention, context-level sequence compression that drops history blocks by target relevance (retention 0.5 costs 0.0001 AUC and saves 75% of compute), 300-candidate sharding, and AOT-compiled serving with no online Python.
One year of Meituan search-ads production data, hundreds of millions of users and items:
| Model | AUC | GAUC | GFLOPs |
| Production base | 0.8558 | 0.8076 | 9.69 |
| RankMixer+DSIN+CIM | 0.8661 | 0.8171 | 171.46 |
| UniCon-Small (0.09B) | 0.8683 | 0.8184 | 201.60 |
| UniCon-Large compressed (0.33B) | 0.8697 | 0.8194 | 197.14 |
UniCon-Small at 0.09B parameters already clears every context-augmented research baseline, so the gain tracks the organization rather than capacity. The most damaging ablation removes hierarchical context modeling entirely (AUC 0.8637). A seven-day full A/B test reports +3.09% RPM, +2.07% CTR, and +2.95% revenue, all p<=0.01. Compression lifts serving throughput 258% over the fixed-length uncompressed baseline.
This is a clean demonstration that input organization can be worth more than model size at fixed parameters, beating backbones with context modules bolted on. For context-rich surfaces, shelves, waterfall feeds, search result lists, the recipe transfers directly. The compression-accuracy curve, 75% of compute saved for 0.0001 AUC, is a reusable template for long-history modeling under serving constraints.
Exposure and position labels come from the incumbent production policy, so supervision regularizes toward the logged display distribution rather than a policy-independent optimum; the authors acknowledge this and it risks entrenching current ranking behavior. Online gains are validated in one search-ads surface; context-sparse feeds may not benefit equally. Whether research baselines received comparable tuning resources on the production data is not stated. Position-conditioned scoring enumerates feasible positions per candidate, and serving cost is only said to fit the SLA. Context units assume complete exposure logs, and robustness to missing impression data is untested.