How UniSAR Unifies Search and Recommendation with Fine‑Grained User Behavior Modeling

TL;DR: The paper introduces UniSAR, a unified framework for modeling fine‑grained user conversion behaviors between search and recommendation, enabling a single service that handles both tasks.

Many platforms now provide both search and recommendation, creating a natural correlation between the two user actions. Existing approaches either model them separately or ignore the different conversion types. The authors identify four conversion patterns—search‑to‑search (s2s), search‑to‑recommendation (s2r), recommendation‑to‑search (r2s), and recommendation‑to‑recommendation (r2r)—as illustrated in the figure below.

Using the KuaiSAR dataset, the authors analyze whether users’ information needs change when moving between search and recommendation. For a randomly sampled set of recommended items, the relevance of the previously clicked item drops from 7.99% (previous scene also recommendation) to 4.86% when the previous scene was search. Conversely, for clicked items in search, relevance falls from 17.14% (previous scene also search) to 3.67% when the previous scene was recommendation. These results indicate that switching contexts often creates new information needs, motivating fine‑grained transition modeling.

UniSAR addresses this by three steps:

Extraction: A transformer with predefined masks extracts features from user actions.

Alignment: Contrastive learning aligns the extracted fine‑grained transition representations.

Fusion: Cross‑attention fuses different transition types into a unified user representation.

The learned representation is fed into downstream search and recommendation models, and the system is jointly trained on both search and recommendation data, allowing knowledge transfer and complementary improvements.

Experiments on two public datasets demonstrate that UniSAR simultaneously enhances search and recommendation performance. Detailed analysis confirms that the model successfully captures user transition behaviors, leading to higher quality recommendations.