Alibaba’s AI Breakthroughs at KDD 2019: From CTR Prediction to Graph Learning

ACM SIGKDD is the premier conference for data mining research, attracting over 3,000 scholars and industry experts each year. In 2019, Alibaba presented more than 20 papers across a range of topics, showcasing the company’s data‑driven innovations.

Paper 1: Practice on Long Sequential User Behavior Modeling for Click‑Through Rate Prediction

To model user interests from long behavior sequences (average length >1000 in Alibaba’s e‑commerce), the authors propose a co‑designed solution combining the Multi‑channel user Interest Memory Network (MIMN) and an independent User Interest Computing (UIC) module. MIMN leverages memory networks and a novel regularizer to capture diverse interests, while UIC decouples interest inference from real‑time CTR estimation, enabling incremental, low‑latency serving.

Paper 2: Personalized Outfit Generation for Fashion Recommendation at Alibaba iFashion

Using a Transformer encoder‑decoder, the model jointly learns visual coherence and personalized recommendation, achieving superior performance in both outfit realism and personalization on large‑scale industrial data.

Paper 3: Personalized Attraction‑Enhanced Sponsored Search with Multi‑task Learning

The authors introduce a multi‑task learning framework that selects personalized selling points for ad titles based on keyword understanding, improving click‑through rates by jointly modeling user preferences for selling points and products.

Paper 4: Exact‑K Recommendation via Maximum‑Clique Optimization

Exact‑K recommendation is formulated as a maximum‑clique problem on a graph. The proposed GraphAttention Network (GAttN) with an encoder‑decoder architecture, trained via Reinforcement Learning from Demonstrations, achieves state‑of‑the‑art results on MovieLens and Taobao datasets.

Paper 5: Sequential Scenario‑Specific Meta Learner for Online Recommendation

A few‑shot meta‑learning approach combines scenario‑specific learning with model‑agnostic sequential meta‑learning, producing a unified framework that quickly adapts to cold‑start recommendation tasks.

Paper 6: A Minimax Game for Instance‑based Selective Transfer Learning

The MGTL model uses a selector, discriminator, and transfer module in an adversarial game to sample source‑domain instances similar to the target domain, improving transfer effectiveness for text matching and e‑commerce recommendation.

Paper 7: IntentGC: A Scalable Graph Convolution Framework Fusing Heterogeneous Information for Recommendation

IntentGC builds a bipartite heterogeneous graph and employs a triplet loss to learn user preferences, integrating multiple information sources with separate graph‑view and node‑view neural networks for efficient large‑scale recommendation.

Paper 8: Metapath‑guided Heterogeneous Graph Neural Network for Intent Recommendation

MEIREC models intent recommendation as a heterogeneous information network, using metapath‑guided neighbor aggregation and a unified word embedding mechanism, achieving a 2.66% uplift in UCTR on the Taobao app.

Paper 9: Isa Single Vector Enough? Exploring Node Polysemy for Network Embedding

The authors propose a multi‑dimensional embedding method that captures multiple aspects of a node, inspired by word polysemy, and demonstrate its effectiveness for classification and link prediction.

Paper 10: Bid Optimization by Multivariable Control in Display Advertising

Using a primal‑dual linear programming formulation, the paper derives optimal bidding strategies under budget and CPC constraints, and introduces a multivariable control system to adapt bids in real‑time bidding environments.

Paper 11: Representation Learning for Attributed Multiplex Heterogeneous Network

A unified framework for embedding multiplex heterogeneous networks is presented, supporting both transductive and inductive learning with theoretical analysis and improved generalization.

Paper 12: Towards Knowledge‑Based Personalized Product Description Generation in E‑commerce

KOBE extends the Transformer encoder‑decoder with knowledge‑aware components to generate informative and personalized product descriptions, incorporating product aspects, user categories, and knowledge bases.

Workshops and Invited Talks

Two workshops were held: (1) Deep Learning Practice for High‑Dimensional Sparse Data, focusing on challenges of applying deep learning to industrial‑scale sparse data; (2) Deep Learning on Graphs, covering graph neural network methods and applications. An invited talk presented AliGraph, Alibaba’s comprehensive graph neural network platform that accelerates graph construction and training by an order of magnitude.