How Alibaba’s ICBU Algorithm Team Transformed E‑Commerce in 2020

Preface

The author habitually writes annual summaries; in 2020 the role shifted from vertical advertising algorithms to a horizontal, overall algorithm portfolio for Alibaba.com International (ICBU).

ICBU Algorithm Team Overview

The team belongs to the ICBU technology division and serves the ICBU business. Its work is organized into four parts: an introductory overview, a discussion of algorithm concepts, an explanation of e‑commerce algorithms, and a detailed look at the ICBU algorithm architecture.

What Is an Algorithm Engineer?

Two perspectives are presented: the "tool" view (engineers who primarily use machine‑learning or optimization techniques) and the "purpose" view (engineers whose main goal is to optimize quantifiable business metrics). Both definitions shape the skill set and mindset of algorithm engineers.

E‑commerce Algorithms

E‑commerce algorithms focus on resource allocation (budget, attention, time, etc.) and the supporting infrastructure needed for effective allocation. The evolution of requirements includes moving from neutral market‑driven allocation to proactive macro‑control, from combinatorial optimization to machine‑learning, from supervised learning to reinforcement learning, from interpretable tree models to deep neural networks, from offline modeling to online real‑time services, and from single‑objective to multi‑objective constrained optimization.

ICBU Algorithm Overview

The team’s workflow is divided into three major components—Understanding, Growth, and Matching—mirroring the marketplace’s "goods, people, place" triad.

Understanding

Builds a digital foundation using computer vision, natural language processing, deep learning, data standardization, and knowledge graphs to improve the digital representation of products, content, buyers, sellers, and market trends.

Growth

Optimizes resource distribution under fixed cost constraints to maximize buyer and seller growth and intelligent operations. It includes buyer growth via combinatorial optimization and forecasting, seller growth via data‑driven modeling and causal inference, and operational growth through optimized allocation of operational incentives.

Matching

Handles efficient matching in search, recommendation, and advertising, aiming to maximize long‑term market efficiency while balancing fairness, commercial versus organic placements, and various stakeholder constraints.

2020 ICBU Algorithm Work Summary

The 2020 achievements are presented under the three pillars.

Understanding

Scene‑material mining for B‑class procurement, generating CPV‑based market segments and simulated purchase scenarios, leading to increased AB and payment buyer metrics.

Intelligent product publishing with a fine‑tuned BART text generation model and ICBU‑specific corpus, improving product title richness by ~6% and adoption rates of 32% (CGS) and 42% (GGS).

Fine‑grained image classification using subject segmentation and graph relational networks, enhancing image tag precision and recall across top industry categories.

Video detection, analysis, and creative generation pipelines that automate multi‑size video synthesis and creative optimization, reducing creative costs.

Growth

Intelligent budget allocation (Version 1.0) using a predictor‑solver architecture and hierarchical reinforcement learning, reducing CPC by 10.3% and publishing novel solutions on self‑attention‑based RL and distributed neuro‑evolution.

Horae ranking system built from scratch for paid traffic, improving AB rates by 13.6% (App) and 3% (Wap).

Supply‑demand matching project (XianZhi) introducing three‑dimensional indices (blue‑sea, competitiveness, richness), shortening order cycles by 8% and boosting MC15 by 44%.

Stellar project predicting high‑quality purchase orders with tree models, raising PO confirmation rates by 7 pp and O‑P conversion by 1.2%.

TAO new‑customer acquisition using SHAP and sub‑models for explainable predictions, increasing conversion by 8.46% and GMV contribution.

Logistics fee sensitivity prediction improving monthly payment buyer growth and ROI.

Matching

Dynamic heterogeneous graph embedding (DyHAN) for evolving buyer‑seller relationships, boosting inquiry conversion by 3.54% and order creation by 14.23%.

Deep Multi‑Interest Network (DMIN) for multi‑category buyer preferences, raising exposure click‑through rate by 10.4% and order conversion by 13%.

Vector‑based retrieval (FashionBERT) combining image patches and text tokens, achieving significant gains on the FashionGen benchmark and accepted at SIGIR 2020.

Semantic search 1.0 (vector recall 3.0 + semantic matching 1.0) improving exposure share by 6%, reducing zero‑result rate by 8%, and lifting click, inquiry, and payment conversion.

Category prediction upgraded from fastText to BERT and enhanced with NER, delivering +5% top‑1 accuracy and +12% recall improvements.

Cross‑language vector recall using EcomLM, reducing zero‑result rate below 1% and improving L‑AB by 1.34% and L‑P by 4.2%.

Reflections and Future Outlook

The author discusses the interplay between data assets and algorithms, emphasizing that data science defines goals while algorithms execute them. Future work will focus on tighter integration of intelligent operations and buyer‑seller growth, redefining optimization objectives for the large‑scale market, and addressing fairness and regulatory constraints.

Recruitment (Removed)

The original text contained a hiring call‑to‑action, which has been omitted as it is promotional.