A robust personalized location recommendation based on ensemble learning

Abstract

Recommender systems (RSs) have attracted considerable attention with the aim of optimizing location service efficiency since a large volume of information is generated by location-based social networks. Prediction accuracy is generally considered the main performance evaluation criterion, while the stability of RSs, which might be affected by uncertainty, has rarely been documented. To guarantee the robustness of RSs on the basis of two essential elements, accuracy and stability, this paper proposes an ensemble-based personalized location recommendation (EPLR) algorithm, in which several different categories of individual models are pre-trained to provide a knowledge base, and the accuracy metric in terms of F1 and the information gain (IG) are individually calculated for each user and are used as personalized weights to integrate the individual models. Notably, IG is used as an evaluation index of system stability. To demonstrate and evaluate EPLR in detail, four representative recommendation algorithms, i.e., user-based collaborative filtering (UBCF), singular value decomposition (SVD), friend-based collaborative filtering (FCF) and kernel density estimation (KDE), are selected as individual models for demonstration. Extensive experiments are conducted on two popular datasets: Brightkite and Gowalla. Additionally, three published ensemble recommendation algorithms and four individual models are implemented for comparison. The experimental results show that EPLR outperforms the other considered algorithms in terms of prediction accuracy and exhibits a promising advantage in system stability.