Guardian : Evaluating Trust in Online Social Networks with Graph - PowerPoint PPT Presentation

Guardian : Evaluating Trust in Online Social Networks with Graph Convolutional Networks Wanyu Lin , Zhaolin Gao, Baochun Li University of Toronto

Almost 4.57 billion people were active internet users as of April 2020. — Statista

Social trust is the basis of online social networks.

Estimates of social trust help indicate to what extent a user could expect someone else to perform given actions, therefore has many applications, such as trust-based recommendations.

B C 2 1 3 3 F A E G 2 1 3 1 1 D H Network graph an example

B C 2 1 3 - I - 3 F A E G 2 1 - i - • 3 1 1 D H Can A trust E? And, to what extent?

Large-scale

Wait a second …

Graph convolutional neural networks — an efficient variant of convolutional neural networks on graphs. 9

� �� � Representation learning with graph convolutional networks Grungier GrghGni_n.FI Output i if lilli __ __ ei __ 10

The complexity of model parameters are independent of the input graph size.

Preliminaries: trust properties B C 2 1 3 3 F A E G 2 1 3 1 1 D H

Asymmetry : one user may trust someone else more than she is trusted back. 3 E G 1 B C 2 1 3 3 F A E G 2 1 Trust properties 3 1 1 D H

Propagative nature : trust may be passed from one user to another, creating chains of social trust that connects two users who are not connected. ? A B C 3 2 B C 2 1 3 3 F A E G 2 1 Trust properties 3 1 1 D H

Composable nature : trust needs to be aggregated if several chains of social trust exit. B 1 3 A E ? 3 1 B C 2 D 1 3 3 F A E G 2 1 Trust properties 3 1 1 D H

An effective way of evaluating trust should be able to capture these trust properties simultaneously. 16

Guardian : an end-to-end learning framework for social trust evaluation. 17

Embedding layer We use a pre-trained embedding layer that maps each user into a vector. I z I 3 I • • s z I B C I - 2 3 B • 8 O 1 E 3 z I 1 : ± g 3 Z I I F A E G 2 1 3 B • g I O z I I 3 3 1 1 I • • I , • I 3 B • g I I D z H I 3 B • 1 8 y I z I 3 18 B • g I

Two types of trust interactions: popularity trust and engagement trust 19

Popularity trust : the overall trust of a user endorsed by others (accumulated from the incoming links) B C 2 1 3 - I - 3 F A E G 2 1 - i - • 3 1 1 D H

Engagement trust : the willingness of a user to trust others (accumulated from the outgoing links) B C 2 1 3 - I - 3 F A E G 2 1 - i - • 3 1 1 D H

B B 1 Popularity Engagement 3 E G E G 3 1 A F A 2 1 1 D H 3 D B C Two types of trust 2 1 3 - I - 3 F A E G aggregation 2 1 - i - • 3 1 1 D H

Trust Relationship Softmax Fully Connected Layer Prediction Layer ? Trust convolutional layer Trustor Trustee Convolutional Concatenation Layers Trust Mean Mean ... ... I 0 I 1 I i O O O 0 1 n Popularity Trust Engagement Trust Propagation Propagation Embedding Layer ? Trustor Trustee To capture the composable and asymmetric nature of trust 23

Trust Relationship Softmax Fully Connected Layer Prediction Stack multiple trust Layer convolutional layers ? Trustor Trustee Convolutional Concatenation Layers Trust Mean Mean ... ... I 0 I 1 I i O O O 0 1 n Popularity Trust Engagement Trust Propagation Propagation Embedding Layer ? Trustor Trustee To capture the propagative nature of trust

Trust Relationship Softmax Fully Connected Layer Prediction Prediction Layer ? layer Trustor Trustee Convolutional Concatenation Layers Trust Mean Mean ... ... I 0 I 1 I i O O O 0 1 n Popularity Trust Engagement Trust Propagation Propagation Embedding Layer ? 25 Trustor Trustee

Trust Relationship Softmax Fully Connected Layer Prediction Layer Guardian ? Trustor Trustee Convolutional Concatenation Layers Trust Mean Mean ... ... I 0 I 1 I i O O O 0 1 n Popularity Trust Engagement Trust Propagation Propagation Embedding Layer ? Trustor Trustee

Our experimental results… 27

Datasets Used Advogato and Pretty-Good-Privacy (PGP) adopt the concept of the “web of trust”, and both contain four different levels of trust. 28

Accuracy Evaluation Accuracy on Advogato 29

Accuracy Evaluation Accuracy on PGP 30

Efficiency 31

Scalability Time vs. # of pairs 32

Scalability Time vs. # of users 33

Guardian is an end-to-end learning framework, that can achieve the best possible performance for social trust evaluation in online social networks.

Wanyu Lin , Zhaolin Gao, Baochun Li University of Toronto 35