A Semi-supervised Social Relationships Inferring Model Based On Mobile Phone Data

With the popularity of mobile phones and the diversification of the sensors in the mobile phone, these sensors producing a large amount of data and the social relationships inferring based on these sensors' data has become a hot research proble m in the field of pattern recognition and ubiquitous computing. The social relations inference models based on mobile sensor data mostly used supervised machine learning methods now and the researchers have an assume that the known amount of social relations labels in the social network is large. But, the relationships is hard and expensive to get in practice. So how to use a small amount of the relationships in the mobile communications network and a large number of sensor data to infer the social relationships is a problem which is badly in need to solution. Through the analysis of the characteristics of the mobile sensor data, this paper puts forward a semi- supervised social relations inference model based on mobile data which use a small amount of label information as well as a large number of cell phone data information to infer social relationships.In mobile communication networks, the biggest problems in relationship prediction model is that the labels are sparse and relationship categories are imbalance. By mining the sensor data and defining the appropriate attributes is one of the methods of improving the precision of the relation inference. Establish a proper relationship inference model can also improve the accuracy of the relation inference. Through the concrete analysis of the sensor data type we defines 96 attributes to infer relationship. In order to solve the problem of relationship label spare and imbalance in the label categories we build a co-training semi-supervised model based on SVM and Naive Bayes. To improve the model we add social balance theory into the relation label infer process.Finally the proposed model was verifyed by MIT laboratory Reality Mining data. Through the analysis of 94 users' mobile data, the model can achive an accuracy of 85% under the condition of known relationship labels is 5% in social network and relationship categories serious imbalance. Compared with supervisd model SVM and naive our model has 20% and 5% improvement respectively. Compared with a semi-supervised model TSVM our model has a 10% improvement.