Multi Interests for Joint Search-Recommendation Modeling
Xiangchen Pan, Wei Wei, Huakang Niu, Zhicong Cheng
cs.IR
2026-08-11
MIJSR mines structural and semantic multi-interests from mixed search-recommend sequences, lifting search HR@5 by about 4.5% on KuaiSAR.
Search and recommendation are two routes into user preference: search is the user actively stating a need, recommendation is passive consumption. A growing line of work models them jointly, feeding a single mixed behavior sequence to serve both tasks.
The trouble is that existing methods encode search and recommendation behaviors into one unified sequence and lose sight of the fact that a user's interests differ by scenario. Someone who browses gadgets and designer toys long-term may suddenly need to book a flight; in a unified model that short-term travel intent gets drowned out by long-standing historical interests.
MIJSR (Multi-Interest Joint Search-Recommendation) pulls the interests apart inside the mixed sequence. Three modules:
On two public datasets, KuaiSAR and Amazon Kindle, against UniSAR and other baselines:
| Task | Dataset | Metric | MIJSR vs baseline |
| Search | KuaiSAR | HR@5 | 0.7811 vs 0.7412 (+4.5%) |
| Search | KuaiSAR | NDCG@5 | 0.6635 vs 0.6397 (+3.7%) |
| Rec | KuaiSAR | HR@5 | 0.5115 vs 0.4887 |
| Search | Kindle | HR@5 | 0.8265 vs 0.8149 |
Search gains are clearer; recommendation also moves up slightly.
For teams building unified search and recommendation, splitting interests by scenario is a sensible, effective direction, especially for preserving short-term intent inside a mixed sequence. This is an incremental improvement, and every module reuses off-the-shelf components (contrastive learning, PLE, clustering), so the bar to deploy is low.
The authors acknowledge a multi-task seesaw: search and recommendation can trade off, one rising as the other falls, and they early-stop on average NDCG@5 to balance it. The cluster count k for semantic clustering is sensitive: too few is coarse, too many is sparse and noisy. The reproduced UniSAR baseline does not match the original paper, which the authors attribute to the seesaw and early-stopping differences, so the absolute gain may not be fully comparable.