Explicable Location Prediction Based on Preference Tensor Model

摘要

Location prediction has been attracting an increasing interest from the data mining community. In real world, however, to provide more targeted and more personal services, the applications like location-aware advertising and route recommendation are interested not only in the predicted location but its explanation as well. In this paper, we investigate the problem of Explicable Location Prediction (ELP) from LBSN data, which is not easy due to the challenges of the complexity of human mobility motivation and data sparsity. In this paper, we propose a Preference Tensor Model (PTM) to address the challenges. The core component of PTM is a preference tensor, each cell of which represents how much a user prefers to a specific place at a specific time point. The explicable location prediction can be made via a retrieval of the preference tensor, and meanwhile a motivation vector is generated as the explanation of the prediction. To model the complicated motivations of human movement, we propose two motivation tensors, a social tensor and a personal tensor, to represent the social cause and the personal cause of human movement. From the motivation tensors, the motivation vector consisting of a social ingredient and a personal ingredient can be produced. To deal with data sparsity, we propose a Social Tensor Decomposition Algorithm (STDA) and a Personal Tensor Decomposition Algorithm (PTDA), which are able to fill missing values of a sparse social tensor and a sparse personal tensor, respectively. Particularly, to achieve a higher accuracy, STDA fuses an additional social constraint with the decomposition. The experiments conducted on real-world datasets verify the proposed model and algorithms.