adversarial learning on heterogeneous information networks

RESEARCH POSTER PRESENTATION DESIGN © 2015

www.PosterPresentations.com

Background & Problem

HeGAN : The Proposed Model

1Beijing University of Posts and Telecommunications2AI Department, Ant Financial Services Group

3Singapore Management University

Binbin Hu1, 2, Yuan Fang3, Chuan Shi1Adversarial Learning on Heterogeneous Information Networks

Heterogeneous Information Network (HIN)l Include multiple types of nodes and linksl Model heterogeneous data and contain rich semantics

HIN Embeddng with GAN based Adversarial Learning (HeGAN)

ExperimentsDatasets

Conclusionsl We are the first to employ adversarial learning for HIN

embedding, in order to utilize the rich semantics on HINsl We propose HeGAN that is not only relation-aware to capture

rich semantics, but also equipped with a generalized generatorl Extensive experimental results have revealed the effectiveness

and efficiency of HeGAN

AcknowledgementsThis research was supported by the National Natural Science Foun- dation ofChina (No. 61772082, 61702296), the National Key Research andDevelopment Program of China (2017YFB0803304), the Beijing MunicipalNatural Science Foundation (4182043) and the Singapore Ministry ofEducation (MOE) Academic Research Fund (AcRF) Tier 1 grant (ApprovalNo. 18-C220-SMU-006).

Contactl hubinbin@bupt.edu.cn

l (bin.hbb@antfin.com)

l shichuan@bupt.edu.cn

l yfang@smu.edu.sg

HIN Embedding (HINE)l Consist of two samplers and one loss functionl Samplers : select positive and negative examplesl Loss function : trained on these samples to optimize

node representations

Adversarial Learning (or GAN)l Makes the model more robust to sparse

or noisy datal Provides better samples to reduce the

labeling requirementl GraphGAN, ANE, NETRA, ARGA

Limitation of HINEl Randomly select existing nodes in the network as

negative samples l Heed to the latent distribution of the nodes so that

lack robustnessl Require domain knowledge

Limitation on GAN based Embeddingl Only investigate homogeneous networks l Poor performance on semantic-rich HINs

Our Idea

HINE+ Adversarial Learning

How to capture the semantics of multiple types of nodes and relations?

How to generate fake samples efficiently and effectively?

Relation-aware, generalized generator

Relation-aware discriminator

Calculate gradient for three cases

I. Connected under given relation

II. Connected under incorrect relation

III. Fake node from relation-aware generator

Challenges

Relation-aware Generator and DiscriminatorI. Discriminator can tell whether a node pair is real

or fake w.r.t relationII. Generator can produce fake node pairs that mimic

real pairs w.r.t relation

Solutions

Generalized GeneratorI. Sample latent nodes from a continuous

distributionII. No softmax computation and fake samples are

not restricted to the existing nodes

Node Clustering

Node Classification

Link Prediction

Adversarial Learning

Heterogeneity and Generalized Generator

EfficiencyRecommendation