QIN: A Query‑Dominated User Interest Network for Personalized Search

Personalized search systems rely heavily on user behavior, yet traditional methods focus only on sparse search actions, ignoring richer historical interactions that could improve relevance and diversity. This paper identifies two main challenges: the mismatch between full‑site behavior and search queries, and the sparsity of search‑only signals.

To address these issues, the authors propose QIN, a query‑dominated user interest network composed of two cascaded modules. The Relevance Search Unit (RSU) first filters the raw behavior sequence for items related to the current query, then further selects the most relevant items for the target, effectively expanding the useful behavior context while mitigating noise.

The Fused Attention Unit (FAU) decouples attention computation for ID features and content attributes, weighting them according to user engagement depth (e.g., fully watched videos receive higher importance). The unit then fuses these attentions, allowing the model to capture varying satisfaction levels across similar items.

Both units rely on pretrained multimodal embeddings to map queries and items into a shared space, using cosine similarity for relevance scoring. The RSU’s two‑stage retrieval first extracts a long‑term interest subsequence for a query, then refines it for a specific target item, providing a richer yet query‑specific behavior representation.

In online serving, QIN adds an RSU module that stores pretrained vectors, indexes them, and retrieves the top‑K most similar historical actions, fitting within real‑time training pipelines. The authors compare QIN against several strong baselines (HEM, ZAM, TEM, GraphSRRL, IHGNN, MultiResAttn) on multiple public datasets, showing substantial improvements across all ranking metrics.

Extensive ablation studies evaluate the contributions of RSU and FAU, sequence length, model size, and hyper‑parameters, confirming the effectiveness of each component. Additional A/B tests on a large‑scale platform further validate QIN’s practical benefits.

Overall, the work demonstrates that integrating full‑history user behavior through query‑driven relevance filtering and decoupled attention yields a more expressive and accurate personalized search model.