Developing a Framework for Next Point-of-interest Recommendation from Spatiotemporal Data

Point-of-interest (POI) recommendation system is popularly used in location based social networks where the goal is to recommend interesting unvisited locations to users. The sequential nature of check-ins hindered many researchers to apply Recurrent Neural Network (RNN) based models for this task. However, most of the models consider only historical check-ins of the user for generating recommendations and fail to incorporate information about current location and time which plays an important role. For reducing data sparsity in spatial dimension, many models use hierarchical gridding of the map which can not reflect spatial distance properly between neighboring grids. Besides, most of the existing models ignored the impact of weather condition when generating recommendation. Keeping these limitations in mind, in this paper we present a framework for point-of-interest recommendation that can model sequential nature of check-ins using Long Short-Term Memory (LSTM) network. We incorporate current spatiotemporal information with weather condition that can provide better personalized recommendation. Instead of hierarchical gridding, we perform linear interpolation for smooth representation of distance between two locations. Extensive experiments on two real world dataset shows that our proposed method surpasses existing state-of-the art methods by 16-18%.