Alibaba’s 5 KDD 2017 Papers Reveal AI, Graph Computing & Large-Scale Learning

On May 23, it was announced that Alibaba Group and Ant Financial had five papers accepted at the 2017 International Conference on Knowledge Discovery and Data Mining (KDD), marking another academic breakthrough for the company after winning the right to host KDD Cup 2017.

KDD (Knowledge Discovery and Data Mining) is the premier international conference on data mining, organized by the ACM Special Interest Group on Data Mining. In 2017 the conference received 1,144 submissions and accepted 216 papers, including 25 from China’s academia and industry.

The five Alibaba papers span deep learning, large‑scale graph computing, intelligent ad ranking, and distributed machine‑learning systems, and many of the proposed methods have already been deployed in Alibaba’s real‑world services.

OCPC Intelligent Pricing Algorithm in Taobao Display Ads

This paper introduces a novel Optimized Cost‑per‑Click (OCPC) algorithm that dynamically adjusts advertisers’ bids for each traffic unit, improving both revenue for advertisers and overall system efficiency while balancing user experience and platform welfare.

Local Algorithm on Large‑Scale Graph Computing for Ad Behavior Prediction

The authors model historical user‑ad interactions as a bipartite graph and propose a random‑walk local algorithm (AdvUserGraph with ADNI) that predicts user behavior with complexity dependent only on the size of the output cluster, achieving effective prediction of rare events in a demand‑side platform.

A New Semantic Encoding Model and Its Application in Intelligent QA and Classification

A convolution‑iterative neural network (conv‑RNN) framework is presented for text semantic modeling, integrating advantages of CNN and RNN structures. The paper also proposes sentence‑level classification and answer‑selection models that outperform state‑of‑the‑art baselines on challenging QA and classification benchmarks.

Multi‑Level Cascade Learning in Large‑Scale E‑commerce Ranking System

To handle billions of items and millions of candidate results under high‑traffic scenarios (e.g., Double‑11), the authors design a cascade search system (CLOES) that uses progressively more complex features across stages, dramatically reducing latency while improving ranking accuracy.

Parameter‑Server Based Distributed Learning System and Its Applications at Alibaba and Ant

Facing TB‑PB scale data and models with billions of parameters, the paper describes “Kunpeng,” a generic distributed platform that combines system‑level and optimization‑level techniques (data/model parallelism, load balancing, fault tolerance) to accelerate algorithms such as LR, FTRL, GBDT, FM, and deep learning on massive datasets.

Alibaba’s e‑commerce business now exceeds 3.7 trillion RMB, generating EB‑scale data that fuels these research efforts, and the company continues to collaborate with top academic institutions to push the frontiers of big‑data and AI technologies.