UnifiedSSR+: An Enhanced Framework for Unified Sequential Search and Recommendation

Abstract

In this work, we propose an enhanced framework named Unified Sequential Search and Recommendation (UnifiedSSR+) for joint modeling of user behavior across both search and recommendation scenarios. Specifically, we consider user-interacted products in the recommendation scenario, user-interacted products and user-issued queries in the search scenario as three distinct types of user behavior. We propose a dual-branch network with a Causal Siamese Encoder to jointly model interacted products and issued queries in search. The query branch is deactivated for recommendation, enabling flexible cross-scenario adaptation. Parameter sharing across branches integrates cross-view and cross-scenario signals. To capture dynamic intent, an Intent-oriented Session Modeling module infers semantic sessions using contextual representations, guided by self-supervised objectives for intra-sequence discrimination and alignment. To further enhance representation learning, the framework introduces multiple attention mechanisms. To emphasize the importance of the current query in capturing real-time user intent, we incorporate a Multi-head Target Attention mechanism, which uses the current query as the retrieval key to selectively aggregate relevant historical queries, improving semantic understanding in the search scenario. Meanwhile, to ensure temporal consistency between training and inference, we introduce a Causal Attention mechanism within the Siamese Encoder, which enforces strict left-to-right information flow aligned with real-world temporal dynamics, thereby reinforcing the temporal consistency of sequence modeling. Extensive experiments conducted on three public datasets demonstrate that UnifiedSSR+ outperforms state-of-the-art methods in both search and recommendation tasks. The source code is publicly available at https://github.com/WItaZhang/UnifiedSSR-plus.