Semi-Supervised Cost-Sensitive Personalized Semantic Place Label Recognition Using Multi-Context Data

Personalized semantic place label recognition is a promising research direction in ubiquitous computing.However,there are some problems existed in current approaches.1)Most of them only took related spatial-temporal data into consideration while neglected other context data(e.g.activity data),which are more significant to reflect the user behaviors during place visiting.2)The current approaches were mostly accuracy-oriented.They aimed to minimized the total recognition error and treated all the misclassification cost loss equally.3)The recognition models based on supervised learning cannot achieve a good performance when the number of labeled place is limited.To address these problems,a multi-context data and ensemble semi-supervised cost-sensitive learning based place personally semantic label recognition model is proposed.Firstly,some context features are constructed from multi context data.Secondly,a semi-supervised ensemble learning framework is utilized to introduce the unlabeled data into multiple classifiers co-training.Meanwhile,exploring the context similarity of places to construct a cost matrix,which is used for cost-sensitive learning to minimize the cost loss.Experimental results show that the proposed approach has its superiority to reduce the error and cost loss.