Breaking Homogenization: Redesigning Multi‑Scenario Recommendation by Uncovering Knowledge‑Sharing Traps

1. Causes and Empirical Analysis

Multi‑scenario recommendation (MDR) shares knowledge across domains to improve per‑scene metrics (CTR/CVR). Two mechanisms cause overall GMV decline despite higher per‑scene accuracy:

Negative transfer from cross‑domain knowledge sharing : shared parameters or features make domain models learn increasingly similar representations, eroding domain‑specific signals.

Greedy “Next Item” optimization : existing MDR methods over‑optimize the immediate next‑item conversion, ignoring long‑term platform health and intensifying competition for limited user attention.

Analysis of Lazada user‑behavior data shows that homogenization rises sharply with the number of page views, reducing item diversity and lowering subsequent purchase rates for highly homogenized users.

2. Solution: Industrial‑Grade Adaptation Design

DPOMR (Direct Preference Optimization from Manual Rules) adapts the DPO paradigm from large language models to recommendation. It aligns model outputs with long‑term, platform‑wide preferences using business‑rule‑derived pairwise preferences instead of costly human‑generated RLHF data.

Manual Rules (MR) Definition

Two core rule families generate “better” vs “worse” pairs from massive behavior logs.

Transaction‑oriented MR Rule 1: Transaction > No‑Transaction Samples with a completed transaction (in any domain) are strictly preferred over non‑transaction samples, ensuring that successful purchases dominate the preference ordering.

Domain‑Collaboration MR

Rule 2/3/4: In‑Domain Interaction > Cross‑Domain Interaction

Within the current domain, a transaction (Rule 2), add‑to‑cart (Rule 3), or click (Rule 4) outranks interactions that occur only in other domains. This re‑weights training labels to prioritize domain‑exclusive demand and prevent all domains from converging on the same hot items.

The rules are highly configurable and have been validated in both offline and online experiments.

Reference Model Redefinition

In recommendation, user behavior evolves rapidly, making an outdated previous‑version reference harmful. DPOMR uses the latest policy model as the initial point and trains a lightweight calibration model with a point‑wise loss to serve as the reference, aligning its predictions (e.g., pCTR/pCVR) with the policy model while preserving timeliness.

Loss Function Design

DPOMR adopts a SimPO‑style Bradley‑Terry objective with a target reward margin Δ. The margin is multiplied by a hyper‑parameter to control calibration strength, preserving reference‑model dependence for stable training while enforcing the desired margin.

3. Offline & Online Evaluation

Evaluation Framework

Because offline metrics cannot fully capture cross‑domain temporal effects, the authors combine joint online A/B tests with custom offline metrics.

Joint A/B testing : Users are consistently bucketed across all recommendation domains (home, detail, cart, etc.) using a unified hash rule, allowing observation of cross‑domain impacts.

Offline metrics

EOHR@K (Domain‑Exclusive Order HitRate at Top K) : Proportion of orders attributed to the domain that supplied the item within the top K recommendations, reflecting the model’s ability to promote domain‑exclusive items.

PvPay_all AUC : Predicts overall platform purchase likelihood across all domains, indicating the model’s capacity to maximize total orders.

Results

DPOMR improves EOHR@5 by roughly 1 % and PvPay_all AUC by 1‑2 % over baseline models, confirming enhanced domain exclusivity while preserving overall accuracy.

In online joint A/B tests, DPOMR raises daily CVR by 0.02 % and order volume by 1.49 %, with larger gains as the method rolls out to more domains. Exposure diversity increases, item‑homogenization probability drops by 2.8 %, and domain exclusivity rates rise by 0.44 % and 0.93 % respectively, demonstrating effective mitigation of the vicious competition trap.

4. Conclusion & Outlook

DPOMR provides a rule‑driven paradigm that tackles negative transfer and greedy‑strategy pitfalls of MDR, enriching its theoretical foundation and delivering measurable business impact. Future work will explore deeper integration of large‑language‑model techniques to further refine cross‑domain user‑preference understanding.