Alibaba Innovative Research (AIR) > Data Security and Privacy Protection
Scalable and Robust Graph Representation Learning

Research Themes

Data Security and Privacy Protection

Background

Graphs have been widely used in industry, such as social network, knowledge graphs, recommendation systems and fraud detection. Recently, Graph neural networks(GNNs) have been achieved success in graph learning tasks such as node classification and link prediction. While a large number of GNNs assumed homophilous settings (where nodes with similar features or same class labels are linked together), which is not a common case in industry.

 

Graphs in Alibaba have the following properties: 1) Heterogeneous, graphs are with multiple types including consumer, seller, item, poi, tag et al. 2) Heterophily, nodes linked together may with different features or labels. 3) Large Scale, with multi-type nodes and multi-relational edges, the real-world graph can have more than 1 billion nodes and 100 billion edges. 4) Dynamic, nodes and edges in graphs change over time.

 

Inspired by pretrained language models, Pretrained Graph Models(PGM) show superior capacity to encoded abundant knowledge from massive labeled and unlabeled graphs, which helps improve computational efficiency for downstream tasks.

 

The main goal of this program is to develop scalable and robust graph representation models that can fit in dynamic graphs with heterophily and heterogeneity, potential methods including graph sampling strategy, graph contrastive learning, link prediction, etc.

Target

1.Graph sampling technique to reduce the graph scale while preserve most graph property

2.Scalable GNN for node classification in Heterophily Graph

3.Graph Pretrained Models for multi-type nodes in Dynamic Heterogeneous Graph.

Related Research Topics

  • Graph Sampling Strategy
  • Graph Structure Learning
  • Graph Robustness
  • Pretrained Graph Models
  • Knowledge Graph based Recommendation

Scan QR code
关注Ali TechnologyWechat Account