Advances in Ranking Algorithms for the "Good Goods" Recommendation Scenario

The article introduces the Good Goods recommendation scenario, where ranking is a critical component for improving recommendation efficiency, and outlines a year‑long series of iterations in long‑sequence modeling, multi‑objective ranking, model structure optimization, loss optimization, and LTR.

Long Sequence Modeling: To handle ultra‑long user interest sequences, the authors propose sub‑sequence extraction combined with attention and multiple mean‑pooling strategies, as well as a category‑retrieval sequence + attention mechanism that filters relevant items before attention modeling.

Category Retrieval Sequence: By using fused category keys to retrieve candidate items and applying target attention, the approach improves precision ranking (AUC+1.1%, PV_GAUC+1.7%, CLICK_GAUC+1.7%) and online metrics such as pctr, uctr, uclick, and dpv.

Original Long Click Sequence Modeling: Four methods—mean pooling, fully‑connected + mean pooling, sub‑sequence extraction + mean pooling, and dynamic routing + target attention—are evaluated, with the sub‑sequence extraction + mean pooling solution selected for both coarse and fine ranking, yielding notable online gains.

Multi‑Objective Ranking: The system shifts focus from click‑through to "grass‑planting" actions (add‑to‑cart, favorite). A multi‑objective model predicts several targets simultaneously, using shared embeddings, gradient blocking, and a mixed fusion formula (α=1, β=2.5, γ=15, δ=15) that balances CTR stability with grass‑planting improvements.

Model Structure Optimizations: Various architectures such as CAN, MMoE, ESMM, and model reconstruction are explored. CAN improves offline CTR AUC by 0.3% and online clicks by 0.87%; MMoE and ESMM provide modest gains across CTR, IPV, and conversion metrics.

Loss Optimizations: Experiments with Focal Loss and GHM Loss show offline AUC improvements (~0.3%) but limited or negative online impact, leading to their exclusion from production.

LTR Layer: To flexibly combine multiple objectives, a Learning‑to‑Rank layer is added, with two approaches: Stacking (leveraging real‑time exposure‑click sequences and achieving offline AUC+1.7% and online pctr+1.52%) and Mixed Sampling (balancing grass‑planting and first‑click samples to mitigate click drop).

Normalization & Feature Engineering: Normalization aligns target score distributions around zero, facilitating weight adjustments in the fusion formula and yielding further online improvements.

The article concludes with a summary of the presented work and outlines future directions, including deeper long‑sequence modeling, universal content representation learning, and continued optimization of ranking structures and LTR methods.