MIMA: Multi-Interest Recommendation via Multi-Positive Exclusive Assignment

1 Introduction

• Multi-interest models represent users with multiple preference vectors, but collapsed interests retrieve overlapping candidates and fail to cover diverse preferences.Existing remedies mainly use auxiliary diversity, routing-balance, or disentanglement constraints outside the primary recommendation objective.Source: §1 · p. 1
• Single-positive training hides concurrent positives from one another, repeatedly selecting a dominant interest while leaving other interests with little distinct supervision.The dominant interest can absorb conflicting gradients from different intents and drift toward an averaged direction.Source: §1 · p. 2
• At inference, scores from differently activated interest channels lack a unified normalization basis, making merged candidate rankings unreliable.A high match from a weakly activated interest may be treated as comparable to a score from a dominant interest.Source: §1 · p. 2
• MIMA groups same-request items, generates complementary interests with a causal decoder, and uses Hungarian matching to assign each positive to a distinct interest.A lightweight routing network estimates user-interest activation probabilities and combines them with item-matching scores for calibrated retrieval.Source: §1 · p. 2
• Across three public datasets and an industrial dataset, MIMA consistently outperforms existing methods, followed by an online A/B test.The stated contributions characterize MIMA as the first multi-positive paradigm for multi-interest recommendation.Source: §1 · p. 2

2.1 Candidate Matching

• Candidate-matching methods progressed from collaborative filtering and matrix factorization to two-tower, sequential, neural-scoring, and graph approaches, but typically retain a single user vector.Single-vector representations make diverse user interests difficult to capture.Source: §2.1 · p. 2

2.2 Multi-Interest Recommendation

• Multi-interest methods use multiple vectors, with MIND and ComiRec advancing interest extraction, but existing approaches leave specialization and cross-channel activation modeling incomplete.Prior collapse remedies include REMI and DisMIR, yet they rely on auxiliary constraints or regularization and do not explicitly model user-interest activation.Source: §2.2 · p. 3

2.3 Multi-Positive Learning

• Multi-positive learning appears in multi-label learning, next-basket recommendation, sequential set prediction, and contrastive learning, but its role in differentiated multi-interest representations remains underexplored.Source: §2.3 · p. 3

3.1 Problem Formulation

• A user is represented by an interest matrix containing K vectors, which support multiple retrievals whose candidate items are merged.The matching stage retrieves a likely subset from the item corpus.Source: §3.1 · p. 3

3.2 Single-Positive Paradigm

• Conventional methods encode a behavior sequence into K interest vectors, then hard-route each single positive to its most aligned interest for scoring.The interest extractor is commonly implemented with attention mechanisms or capsule networks.Source: §3.2 · p. 3
• Because concurrent positives are split across instances, they cannot compete for distinct interests; updates favor the selected, often dominant, interest and leave others scarcely supervised.Under this training setup, redundant interest representations can become a stable solution.Source: §3.2 · p. 3

4.1 Multi-Positive Target Construction

• Items clicked in one request form concurrent positive sets, paired with the user’s history available before that request.Source: §4.1 · pp. 3–4

4.2 Complementary Interest Generation

• A causal decoder conditions later queries on earlier interest states, directing them toward uncaptured behavior patterns and producing complementary interests.Source: §4.2 · pp. 4–5

4.3 Exclusive Interest Assignment

• Hungarian matching assigns each valid positive to one distinct interest, making differentiation part of the training objective rather than an auxiliary constraint.Source: §4.3 · p. 5

4.4 User-Interest Routing

• A routing query summarizes the interest queries to estimate activation probabilities, which calibrate per-channel scores for joint ranking.Source: §4.4 · p. 5

4.5 Optimization

• Training combines an interest-item loss for representations and a user-interest loss for routing, with stop-gradient keeping their parameter updates decoupled.Source: §4.5 · p. 6

5.1 Experimental Settings

• Experiments cover Books and Beauty, Gowalla, and Lazada Thailand; public positive sets approximate requests with one-day windows, while Industry uses real serving requests.Public data use 8:1:1 user splits and Recall, NDCG, and HR at 20 and 50; Industry trains on 10 days and tests on the final day using HR at 100, 500, and 1000.Source: §5.1.1 · p. 6, §5.1.3 · p. 7

5.2 Overall Performance Comparison

• MIMA is best across all three public datasets, beating the second-best method by 3.04%–15.72% (p<0.01); gains widen as interaction density falls.Relative improvements range from 3.04%–7.25% on Gowalla, 4.76%–11.64% on Beauty, and 9.38%–15.72% on Books.Source: §5.2 · p. 7, Table 2 · p. 8

5.3 Ablation Study

• Ablations show the largest performance drop without multi-positive assignment; removing routing or the causal decoder also reduces performance.The authors attribute the gains to differentiated training signals, score calibration across interest channels, and complementary interest generation.Source: §5.3 · p. 7, Table 3 · p. 8

5.4 Advanced Analysis

• Hungarian assignment performs best across the three datasets; greedy matching is slightly weaker, while soft Sinkhorn is especially weak on sparse Books.The paper attributes the difference to Hungarian's exact total-similarity optimization, greedy's locally chosen pairs, and Sinkhorn's spread of each positive's gradient across interests.Source: §5.4.1 · p. 7, Table 4 · p. 8
• Best K is 8 on Books and Beauty and 4 on Gowalla; two decoder layers are best or near-best, with optimal heads differing by dataset.Beauty performs best with 8 heads, while Books performs best with 2; increasing K also increases the number of positives used for supervision.Source: §5.4.3 · p. 8, §5.4.4 · p. 8, Figure 4 · p. 9
• In the t-SNE comparison, MIMA's interests are most differentiated, align with distinct behavior clusters, and better cover regions containing future items.The visualization samples two Gowalla users with 150 interactions each and projects historical items, future items, and four interests together.Source: §5.4.2 · pp. 7–8, Figure 3 · p. 9

5.5 Industrial Results

• On Industry, multi-interest methods generally outperform single-interest models, but public-dataset leader DisMIR falls behind; MIMA achieves the highest IDM.The industrial comparison sets the number of interests to 5 for all multi-interest methods.Source: §5.5.1 · p. 9, Table 5 · p. 9
• In the 7-day online test, transactions rose 5.60% and transaction amount 5.44%, while exclusive impressions increased 5.12 points; a sampled user’s five interests map to distinct categories.Each row of the case-study figure pairs historical items for one interest with items retrieved by that interest vector.Source: §5.5.2 · p. 9, Table 6 · p. 9, §5.5.3 · p. 10

• MIMA consistently outperforms state-of-the-art baselines across three public datasets and an industrial dataset, with significant business gains in an online A/B test.Source: §6 · p. 10
• MIMA groups concurrent behaviors into positive sets and exclusively matches each positive to a distinct interest, letting differentiation emerge from the training objective.It generates complementary interests with a causal decoder rather than relying on auxiliary regularization.Source: §6 · p. 10
• A lightweight user-interest routing module estimates each interest’s activation strength to yield calibrated, comparable scores across channels at inference.Source: §6 · p. 10