Modern industrial ranking systems are increasingly adopting generative recommendation (GR) paradigms, where user behavior sequences and non‑sequential (NS) features are tokenized and jointly modeled by a Transformer‑style backbone. This unified formulation handles long histories and heterogeneous features, yet sequence and NS representations often carry asymmetric semantics: historical tokens mix environment context with observed‑action semantics, while candidate‑side NS tokens lack observed actions but contain richer supplementary context such as user‑item interaction features and recent behavior lists. Direct query‑key dot products between these mismatched representations compare different semantic objects, limiting meaningful cross‑token interaction. To address this gap, we propose EASformer, an Environment‑Action‑State disentangled architecture for industrial ranking. EASformer builds an explicitly aligned environment space shared by historical and candidate NS tokens, while disentangling user actions from historical tokens and preserving candidate‑specific context in NS tokens. Leveraging this aligned space, the model enables complementary historical action retrieval and state‑conditioned action prediction. The disentangled design also supports a Cloze‑style self‑supervised objective by masking historical actions and reusing the candidate prediction pathway. Offline benchmarks on a real‑world industrial app marketplace achieve an AUC of 0.89411, a 0.88% relative improvement over the deployed baseline. Online A/B tests show eCPM lifts of 1.26% and 3.94% in two ranking scenarios.
Review