Machine learning techniques for mining location-based social networks for business predictions

Abstract

The vast use of Location-Based Social Networks over the last decade results in a considerably large amount of data transactions accumulated over time. Data mining researchers exploit these large amounts of data produced by location-based social networks users to predict useful information. One of the highly recommended methods for prediction is learning from statistical observations through machine learning. In this paper, spatial machine learning techniques are suggested for predicting potential business investment openings in Location-Based Social Networks. Two spatial techniques are studied; Spatial Auto Regression model, and Inverse Distance Weight spatial interpolation technique. A modification for the Inverse Distance Weight spatial interpolation has been suggested where an extreme avoidance criterion is proposed, in order to enhance the prediction accuracy of the classical method. The modified interpolation method is compared to the spatial Auto Regression model and the classical Inverse Distance Weight spatial Interpolation technique, for considerations of computational complexity and prediction accuracy. An experimental case study involving data extracted from Foursquare has been tested for Business venue usage prediction. The results show better prediction accuracy for the Inverse Distance Weight spatial interpolation than the Spatial Moreover, better prediction accuracy for the proposed Extreme Avoidance spatial interpolation over both the Inverse Distance Weight spatial interpolation and Spatial Auto Regression is obtained while maintaining the low computational complexity of the spatial interpolation.