Context‑Aware Re‑ranking in Industrial Recommendation Systems: Design and Practice of a List Retrieval System

Re‑ranking has become an essential component of industrial recommendation systems, yet traditional greedy approaches ignore the contextual relationships among items in the final list, leading to sub‑optimal sequence rewards.

The authors analyze existing re‑ranking methods, highlight four major challenges (context awareness, permutation specificity, complexity, and business requirements), and argue that a truly optimal solution must model both upstream and downstream influences of each item.

To address these issues, they introduce a two‑stage List Retrieval System (LRS): • Re‑rank‑recall generates candidate sequences using a fast beam‑search algorithm and a Generative Reranking Network (GRN) that leverages policy‑gradient training. • Re‑rank‑sorting scores each candidate with a Deep Context‑Wise Network (DCWN) and selects the sequence with the highest List Reward (LR) metric.

Extensive online experiments on Taobao’s homepage, micro‑detail pages, and short‑video feeds demonstrate notable improvements (e.g., PV + 11%, IPV + 6%, effective VV + 5%, and increased average watch time) with minimal latency overhead.

The paper also discusses practical deployment considerations such as edge‑computing for GRN inference, handling variable sequence lengths, and integrating business‑driven diversification rules.