FriendTransfer: Cold-start Friend Recommendation with Cross-platform Transfer Learning of Social Knowledge

 

Ming Yan, Jitao Sang, Tao Mei, Changsheng Xu

 

 

 


Summary


 

(1)  The proliferation of various social media sites has provided opportunities to cross-platform research and applications.
(2)  Social relation and behavior analysis are investigated in a cross-platform scenario between Twitter and Flickr.
(3)  In cross-platform scenario, typical social relation types include: no relation, unidirectional relation, bidirectional relation and cross-follow relation(A follows B in both platforms).
(4)  We conduct an in-depth data analysis to examine what information can better transfer from one platform to another.
(5)  Based on the analysis, we further design a cold-start friend recommendation application by leveraging the convinced social information.

 


Data Collection


 

 

 

By leveraging user self-provided account links on Google+, we construct a user dataset with user account linkage between Twitter and Flickr. In our final 1,457 user dataset, every user has both accounts both in Twitter and Flickr and they are not isolated.
The users' rich social context information are also downloaded from their Flickr and Twitter accounts, respectively:
(1) 
In Flickr, every user's contact list, interested group list as well as the image set with tags are downloaded.
(2)  In Twitter, each user's friend list and follower list are obtained.

 


Data Analysis


 

Social Relation Analysis

 

Two questions are mainly considered:
(1)  Can social relation transfer across different platforms?
(2)  What type of social relation is easier for transferring?

 

Observation 1. High transfer ratio and more closely connected in Flickr

 

 

Observation 2. Bidirectional relation is more reliable for transferring

 

 

Social Behavior Analysis

 

Two questions are mainly considered:
(1)  Whether some sort of consistency exists between social behavior and relation?
(2)  What type of social behavior is easier for transferring?
(In the tables below, CCN means common contact number, CGN means common interested group number, TBS means tag-based similarity among user pairs.)

 

Observation 1. Social behavior and social relation have some sort of consistency

 

 

 

Observation 2. Common contact and tag-based profile can promote cross-follow relations

 

 


Applications


 

Cold-start friend recommendation

 

Goal

To recommend the top-k friends for experienced users in Flickr who just come to a new social platform Twitter.

 

Model Formulation

We formulate the top-k friend recommendation solution as a random walk with restart over the user graph where users are nodes and the edges between them are weighted by their similarity on social behavior.

 

image004.bmp

 

Based on the social relation and behavior analysis above, we further formulate the element of initial relevance vector and transition probability matrix as below:

 

image004.bmp

where sij denotes the pairwise user similarity computed by their convinced social behavior.

 

 

Experimental Results

Evaluation results on our cold-start friend recommendation application are shown in the figure below.
(1) It demonstrates the advantage of combining social behavior information and social relation information over that only utilize one kind of social information.
(2) It validates the effectiveness of our random walk fusion method over the simple coarse combination method.
(3) The performance is not sensitive to the change of trade-off parameter \alpha which validates the robustness of our algorithm.

 

image004.bmp

 


Publication


 

FriendTransfer: Cold-start Friend Recommendation with Cross-platform Transfer Learning of Social Knowledge [pdf] [slides] [data]

Ming Yan, Jitao Sang, Tao Mei and Changsheng Xu
In IEEE International Conference on Multimedia & Expo (ICME), San Jose, USA, 2013.