A Scalable Knowledge Graph Embedding Model for Next Point-of-Interest Recommendation in Tallinn City

Abstract

With the rapid growth of location-based social networks (LBSNs), the task of next Point Of Interest (POI) recommendation has become a trending research topic as it provides key information for users to explore unknown places. However, most of the state-of-the-art next POI recommendation systems came short to consider the multiple heterogeneous factors of both POIs and users to recommend the next targeted location. Furthermore, the cold-start problem is one of the most thriving challenges in traditional recommender systems. In this paper, we introduce a new Scalable Knowledge Graph Embedding Model for the next POI recommendation problem called Skgem. The main originality of the latter is that it relies on a neural network-based embedding method (node2vec) that aims to automatically learn low-dimensional node representations to formulate and incorporate all heterogeneous factors into one contextual directed graph. Moreover, it provides various POIs recommendation groups for cold-start users, e.g., nearby, by time, by tag, etc. Experiments, carried out on a location-based social network (Flickr) dataset collected in the city of Tallinn (Estonia), demonstrate that our approach achieves better results and sharply outperforms the baseline methods. Source code is publicly available at: https://github.com/Ounoughi-Chahinez/SKGEM