Toward Relational Learning with Misinformation Liang Wu * , Jundong - - PowerPoint PPT Presentation

Arizona State University Data Mining and Machine Learning Lab

Toward Relational Learning with Misinformation

Liang Wu*, Jundong Li*, Fred Morstatter+, Huan Liu*

*Arizona State University +University of Southern California

{wuliang, jundongl, huanliu}@asu.edu, morstatt@usc.edu

Arizona State University Data Mining and Machine Learning Lab

Classification in Social Media

• Relational learning aims to classify linked

nodes in a graph (social networks)

• Task: Classification
• Feature: Attributes, Links

Arizona State University Data Mining and Machine Learning Lab

Classification in Social Media: Our Task

• Relational learning aims to classify linked

nodes in a graph (social networks)

• Task: Classification
• Feature: Attributes, Links
• Challenge: Data is Inaccurate

Arizona State University Data Mining and Machine Learning Lab

Social Media Data is Inaccurate and Noisy

• Attacks of content polluters

– Node attributes cannot reveal the identity

• Colloquial language of regular users

– Misinformation, inaccurate data

Arizona State University Data Mining and Machine Learning Lab

Classification with Noisy Data

• Weighting Nodes
• Anomalous points are lower weighted

– Larger loss leads to smaller weights

Weighted Learning Node Weights

Classifier

?

Arizona State University Data Mining and Machine Learning Lab

Classification with Noisy Social Media Data

• Attacks of content polluters

– Node attributes cannot reveal the identity

• Colloquial language of regular users

– Misinformation, inaccurate data

Arizona State University Data Mining and Machine Learning Lab

Robust Classification with Network Information

• Weighting Nodes with Centrality
• Authoritative points are higher weighted

– – Larger centrality leads to higher weights

Weighted Learning Node Weights

Classifier

?

Arizona State University Data Mining and Machine Learning Lab

Denoising with Social Networks?

• Links can be noisy
• Obtaining all links (complete graph) is difficult

Arizona State University Data Mining and Machine Learning Lab

Community Structures are More Robust

Malicious User

Arizona State University Data Mining and Machine Learning Lab

Community Structures are More Robust

Malicious User Community Detection Malicious User

Arizona State University Data Mining and Machine Learning Lab

Denoise with Community Structures

Arizona State University Data Mining and Machine Learning Lab

Community Candidate Generation + Community Selection

Arizona State University Data Mining and Machine Learning Lab

Community Candidate Generation + Community Selection

𝐱,𝐝 𝐧𝐣𝐨 ෍ 𝒋=𝟐 𝑶

ci 𝐲𝐣𝐱 − 𝐳𝒋

𝟑 + λ1||w||2

2 𝐓𝐯𝐜𝐤𝐟𝐝𝐮 𝐮𝐩 ෍

𝒋

𝒅𝒋 = 𝑳

+ λ2σi=0

d

σj=1

ni ||𝐝Gj

i||2

𝑕𝑠𝑝𝑣𝑞 𝑀𝑏𝑡𝑡𝑝 𝑏𝑤𝑝𝑗𝑒 𝑝𝑤𝑓𝑠𝑔𝑗𝑢𝑢𝑗𝑜𝑕

L1

1 nor

norm on

• n th

the in inter-group p le level L2 norm on

• n th

the in intra-group le level

d: depth of hierarchy of Louvain method ni: number of groups on layer i 𝐝Gj

i: nodes of group j on layer i

Arizona State University Data Mining and Machine Learning Lab

Optimization

𝐱 𝐧𝐣𝐨 ෍ 𝒋=𝟐 𝒏

ci 𝐲𝐣𝐱 − 𝐳𝐣 𝟑 + λ1||w||2

2

Optimize w 𝐱,𝐝 𝐧𝐣𝐨 ෍ 𝒋=𝟐 𝒏

ci 𝑢𝒋

𝐓𝐯𝐜𝐤𝐟𝐝𝐮 𝐮𝐩 ෍

𝒋

𝒅𝒋 = 𝟐

+ λ2σi=0

d

σj=1

ni ||𝐝Gj

i||2

Optimize c

Arizona State University Data Mining and Machine Learning Lab

Evaluation

Results

Macro- and Micro-average of F1-measures with increasing ratio of misinformation

Flickr

Arizona State University Data Mining and Machine Learning Lab

More Results

BlogCatalog

Effectiveness of identifying mislabeled instances

BlogCatalog Flickr

Arizona State University Data Mining and Machine Learning Lab

Conclusions

• A supervised learning method with inaccurate

networked data

– Focusing on community structures instead of links – Can be integrated to other algorithms – Efficient to solve