Improving Bandit-Based Recommendations with Spatial Context Reasoning: An Online Evaluation

The pervasive deployment of low cost WiFi access points has accelerated the development of mobile computing to provide ubiquitous computing. Herein, we focus first on the discovery of urban districts, in several french cities, using the connection history of mobile users to a city-wide free public Wi-Fi network. The goal of our approach is to infer relevant spatial context features that can be exploitable by bandit-based recommendation systems and improve their performances. For the unsupervised context reasoning step, we use spectral clustering to deduce areas by grouping Wi-Fi access points according to their users' visitations. We have published an anonymized sample of our dataset and our results on the web. Then, we have integrated the deduced spatial context into a mobile cultural events visualization and recommendation app in order to evaluate the global performance online. Thus, over a year we have observed how such spatial context improves bandit-based recommendations in this app by comparing two use cases of the LinUCB algorithm: the first using the original context without the deduced geo-context, and the second using context enriched by our computed spatial context. Finally, our online evaluation shows that better results are obtained when combining spatial context reasoning with the bandit-based recommendation system, both in terms of accuracy and user participation.