Data augmentation for next point-of-interest recommendation

Abstract

The problem of where to go next has been highly studied and coined as next or successive Point-Of-Interest (POI) recommendation. In essence, using historical data and drawing contextual inferences, an algorithm or model may be able to predict where a user may be interested to go next. These data are popularly collected on Location Based Social Networks (LBSN) as users’ check-in their location to share their journey with friends. However, in Recurrent Neural Network (RNN) and sequential-based models, the raw data from check-ins in LBSN lead to sparse user sequences (or consecutive journey). To overcome this problem, we dive deeper into the data preparation phase to better understand how data is prepared for sequential models and the sequence generation process. We then propose a new model that applies an ensemble technique. The proposed model contains two important modules, (1) a data augmentation module that helps to generate new artificial check-ins to solve the problem of data sparsity, followed by (2) a sequential behaviour encoder, resulting in better model data input for more contextual and relevant predictions. We perform multiple experiments and our results show significant improvements as compared to related works in deep learning.