SetMIR: Multi-Interest Retrieval as Set Prediction

简体中文

Core mechanism on public data: interest set, presence, and query-level NMS.

Paper information

Field	Value
Paper	arXiv 2608.30251
Affiliations	Snap Inc.
First public release	2026-08-31 (arXiv v1)
Authors' code	No public author implementation found as of 2026-09-05.
Adapter / method	setmir
Local code	src/auto_research/reproductions/setmir/

Original paper

Background and contribution

Decode unordered interests with K learned queries. Hungarian one-to-one matching discourages collapse; presence identifies required interests and query NMS suppresses duplicate ANN requests.

flowchart LR
  H[History] --> E[Transformer encoder] --> D[Set decoder]
  Q[Learned queries] --> D --> P[Presence] --> N[Query NMS] --> A[Dynamic ANN]

Key figure from the paper

Original Figure 1: SetMIR. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism

The set loss is \(\min_{\sigma\in S_K}\sum_k \mathcal L(q_k,y_{\sigma(k)})\). Dispatch only queries passing presence and NMS.

Paper results

The paper reports 33% fewer ANN requests and +3.1% CVR on Snap DPA, with a production retrieval source.

Local reproduction

Public genre clusters proxy interest queries; presence and similarity NMS run. NDCG@10 0.05401→0.05665 (+4.90%), with three-seed metrics.

Reproduction limits

No private DPA history, production Transformer, or ANN dispatch. CPU public-data path only.