Ranking Learning in Mobile Taobao: Challenges, Solutions, and Improvements

Speaker Zhou Liang, a machine‑learning expert from the Chinese Academy of Sciences, introduces his background in large‑scale parallel algorithm optimization, advertising CTR estimation, and the development of the mobile Taobao recommendation platform.

Ranking learning is a core problem for recommendation, search, and advertising; in the constrained display space of Mobile Taobao it becomes especially critical for selecting the most appealing items from billions of products.

Mobile Taobao Recommendation Overview

Figure 1: Full coverage of Mobile Taobao recommendation business.

The system aims to improve user experience with personalized feeds, provide merchants with traffic, and guide platform behavior.

Figure 2: Architecture of the Mobile Taobao recommendation system.

Why Ranking Learning?

Figure 3: Reasons for applying ranking learning.

Ranking learning can be categorized into:

Pointwise (predict a score for each pair)

Pairwise (optimize relative order of item pairs)

Listwise (directly optimize the whole ranked list, e.g., NDCG)

Business Example – In‑shop recommendation: only items from the same shop are shown, the goal is CTR, and the method used is Pointwise.

Figure 4: In‑shop recommendation business.

Sample construction aims to predict CTR and rank by the predicted score. Positive/negative sample handling includes various ratios of clicks, exposures, and conversion adjustments.

Feature Design

Features include ID‑type (user, item, context), mobile‑specific attributes (device ID, city, phone model), and cross features such as gender‑match, age‑match, purchase‑power match, and position bias.

Figure 6: Age‑matching feature engineering.

Real‑time user features capture session information such as categories viewed, previous exposures, and cross‑shop purchases. Online learning is performed with offline feature extraction followed by FTRL updates.

Figure 8: Online learning pipeline.

Industry Market Applications

Figure 9: Personalized module ranking across industry verticals, aiming to maximize clicks while balancing traffic.

Methods include AUC optimization, Pairwise‑RankNet, and handling of position bias.

BPR – Bayesian Personalized Ranking

Key is constructing pair samples such as Click > Skip, Last Click > Skip, Click > Earlier Click, Click > No‑Click Next.

Figure 11: Female‑clothing waterfall flow example.

Multi‑objective waterfall flow optimizes CTR, CVR, and average order value using Listwise‑LambdaMart.

Figure 12: Multi‑objective fusion.

Optimizing NDCG

DCG and its normalized version NDCG are used as evaluation metrics; LambdaRank directly computes gradients to optimize these IR metrics, and LambdaMart combines MART with Lambda gradients.

Figure 13: NDCG optimization illustration.

Feature representation for LambdaMart includes continuous features, feedback dimensions for user/item/context, session features, sub‑model scores, and LBS feedback.

Figure 14: Sample construction for multi‑objective Listwise training.

Future work includes handling massive log collection across many devices, aligning PC and mobile features, expanding LTR applications, and improving real‑time model updates and user behavior mining.

Figure 15: Plan and outlook.