简体中文 · English

MIMA: Learning complementary interests through exclusive multi-positive assignment

Public-data core mechanism. Groups multiple positives, runs minimum-cost one-to-one assignment, updates complementary interests and calibrates activation. Public sequence features replace the production causal Transformer.

Paper information

Field	Details
Paper	arXiv v1
Institution	Alibaba International Digital Commerce Group
First public release	2026-09-11 (arXiv v1)
Original source code	No public author implementation found on 2026-09-15.
Adapter	mima
Local reproduction code	src/auto_research/reproductions/mima/

Original paper summary

Background and main changes

Single-positive training in multi-interest retrieval often updates the same closest interest, causing collapse. MIMA groups co-occurring items from one request as positives, generates interests with a causal decoder, and uses Hungarian matching to enforce one-to-one supervision. Routing estimates each interest's activation probability, making channel scores comparable.

flowchart LR
  H[History] --> D[Causal interest decoder]
  D --> I[Complementary interests]
  P[Same-request positives] --> M[Exclusive Hungarian matching]
  I --> M
  I --> R[Interest activation routing]
  M --> L[One-to-one training objective]
  R --> S[Calibrated multi-channel retrieval]

Key figure from the paper

Figure 2: Core architecture and connections. From the paper; copyright belongs to the authors. Click for the source.

Core equations

For interest \(i_k\) and positive \(p_j\), use cost \(C_{kj}=-i_k^\top p_j\) and solve \(\pi^*=\arg\min_\pi\sum_k C_{k,\pi(k)}\). At inference calibrate interest-channel scores by \(\log a_k\). The mini-suite exhaustively solves small matrices for the exact one-to-one optimum of the same objective as Hungarian matching.

Offline and online results

Industrial offline HR@100/500/1000 improved by 7.98%/12.24%/13.03% relative to the best baseline. Seven-day production A/B reported +5.60% transactions, +5.44% transaction value and a 5.12-percentage-point rise in exclusive exposures (Section 5.5, Tables 5–6).

Local reproduction

See metrics/movielens-100k-seeds42-44.json for MovieLens 100K seeds 42/43/44. Training positives come only from history, with no validation/test answers read. Only validation selects the blend.

Comparison protocol: Baseline NDCG@10=0.05401 versus 0.04589 for the experimental arm, a relative change of -15.03%.

Scope

Public data lacks production same-request multi-purchase logs; nearby events substitute. The private Alibaba catalog, production causal Transformer checkpoint and online retrieval stack are not reproduced. No Evolve mapping is claimed without an executable candidate protocol.