Modern large-scale e-commerce recommendation systems rely heavily on sequence modeling to infer user intent from historical actions. However, traditional models typically treat user journeys as flat lists of item identifiers and summarize user history early, without prior conditioning on the target item. This simplification discards rich contextual cues and limits cross-catalog product discovery, resulting in diluted, suboptimal product recommendations. We introduce SHOPPER (Semantic, Historical, Order-aware, and Product-conditioned Pooling for E-commerce Recommendation) which addresses these bottlenecks through: (a) Semantic history, leveraging an upstream foundation model to extract rich multimodal event embedding of users shopping intent; (b) Order-aware local intent modeling, which enriches historical events with dense temporal metadata (e.g., dwell time, timestamp deltas) and applies a 1D convolutional layer to capture short-term, localized shopping intent bursts; and (c) Product-conditioned pooling, with serving-aware factorization, which splits computation into a heavy request-level phase (event enrichment) and a per-candidate phase (conditioning gate + cross-attention), enabling deep target-aware summarization — effectively isolating candidate-specific signals from noisy browsing histories.
Offline evaluation shows substantial NE reductions, and a live A/B test confirms +0.4% lift in top-line business metric while remaining deployable within strict latency constraints