Authors:
Yuka Ikebe
1
;
Masaji Katagiri
2
and
Haruo Takemura
3
Affiliations:
1
NTT DOCOMO and Inc., Japan
;
2
NTT DOCOMO, Inc. and Osaka University, Japan
;
3
Osaka University, Japan
Keyword(s):
Link Prediction, Matrix Factorization, Latent Feature, Semi-supervised Learning, One-class Setting.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Information Extraction
;
Knowledge Discovery and Information Retrieval
;
Knowledge-Based Systems
;
Mining High-Dimensional Data
;
Symbolic Systems
;
User Profiling and Recommender Systems
Abstract:
This paper describes a semi-supervised learning method that uses smartphone usage data to identify friendship in the one-class setting. The method is based on the assumption that friends share some interests and their smartphone usage reflects this. The authors combine a supervised link prediction method with matrix factorization which incorporates latent features acquired from the application usage and Internet access. The latent features are optimized jointly with the process of link prediction. Moreover, the method employs the sigmoidal function to estimate user affinities from the polarized latent user features. To validate the method, fifty university students volunteered to have their smartphone usage monitored for 6 months. The results of this empirical study show that the proposal offers higher friendship prediction accuracy than state-of-the-art link prediction methods.