Evolution of the First-Focus Personalized Recommendation Model in E-commerce

This article presents a comprehensive case study of how the "first‑focus" (首焦) recommendation slot in an e‑commerce platform was continuously optimized using deep‑learning techniques. The first‑focus slot is the top‑most visual element that instantly conveys product and event information, making it critical for both traffic distribution and user engagement.

1. Background

The slot serves as a primary traffic‑distribution channel. Optimizing its algorithm directly improves click‑through rate (CTR), conversion, and overall user experience. The authors describe a series of algorithmic upgrades that were repeatedly deployed across multiple resource positions, yielding significant gains in user clicks and conversions.

2. System Overview

The recommendation ecosystem consists of data collection, a ranking model that scores candidate materials, and feedback loops that influence both the ranking system and content creators. Position bias, user behavior, and material quality are all considered.

3. First‑Stage: DCN‑mix Single‑Objective Model

Introduced a race‑model (赛马模型) with reinforcement learning to balance exploration and exploitation.

Added a personalized model using T+1 data to capture user interests.

Adopted the DCN‑mix (Deep & Cross Network) architecture, which augments the wide component of a WDL model with a cross network for automatic high‑order feature interactions.

Cross‑Network formula (shown in Figure 2) treats the original input as its own cross‑object at the first layer, similar to residual connections.

Both serial and parallel structures were evaluated; the serial version performed best for the first‑focus scenario.

Online deployment yielded a 4 % CTR lift and a 22.2 % increase in UV value.

4. Second‑Stage: Incorporating BST for Dynamic Feature Capture

Behavior Sequence Transformer (BST) was used to model users' real‑time click sequences, providing richer dynamic features.

BST replaces sparse user‑ID embeddings with dense sequence representations, improving training speed.

Experiments showed BST alone outperformed combinations of user‑ID and BST.

Mapping real‑time item clicks to product categories allowed the model to share embeddings between materials and target items.

Result: +3.24 % exposure‑CTR, +8.33 % exposure‑conversion, +7.69 % UV value.

5. Third‑Stage: Position‑Bias Debiasing

Analyzed the Matthew‑effect caused by higher‑position items receiving disproportionate clicks.

Reviewed industry solutions (Meituan 2017, Huawei PAL 2019, YouTube‑net 2019) and their limitations.

Implemented a shallow tower to predict position bias and combined it with the main ranking model.

Online A/B tests showed a 5.06 % CTR increase and consistent gains in conversion and UV value.

6. Fourth‑Stage: Multi‑Objective Model for Click Efficiency

Added a second objective: average number of products viewed per session (session_len) after a click.

Used a Multi‑gate Mixture‑of‑Experts (MMoE) architecture to balance CTR (classification) and session_len (regression).

Initial experiments suffered from loss‑scale mismatch; later tuning (gradient clipping, DWA weighting) aligned the tasks.

Final formulation converted the regression task to a classification task based on percentile thresholds, enabling ESMM‑style joint training.

Achieved stable CTR while significantly increasing per‑session product views, leading to higher conversion rates.

7. Fifth‑Stage: BST + Target‑Attention for Key‑Product Promotion

Addressed the need to boost exposure for newly launched or promotional items.

Combined BST with a target‑attention (DIN) layer to capture item‑to‑item relevance.

Offline tests showed a 2 % lift in both CTR and UV value; online results gave an 18 % increase in exposure PV for priority items and a 16 % rise in site‑wide product‑detail page UV.

These techniques have been propagated to other recommendation slots within the platform, confirming their general applicability.

8. Conclusion

Stage 1: Sample‑weighting by dwell time and real‑time conversion weighting yielded quick, noticeable gains.

Stage 2: BST introduced dynamic user interest modeling, replacing sparse ID features.

Stage 3: Position‑bias correction mitigated the Matthew‑effect and improved fairness.

Stage 4: Multi‑task learning balanced click efficiency with downstream browsing behavior.

Stage 5: BST + target‑attention further enhanced key‑product exposure without harming overall slot performance.

The iterative methodology demonstrates how practical business constraints and systematic A/B testing can guide deep‑learning model evolution in large‑scale e‑commerce recommendation systems.