Towards an approach for online evaluation of new variants of content-based POI recommender systems by mobile tourists

Recommendation systems (RS) play an important role in our everyday life, especially in domains such as e-commerce, online advertising and tourism. Indeed, these RSs provide suggestions that are closer to user expectations, thanks to algorithms inspired by user previous behavior. In the literature, content-based RS are used for the recommendation of movies, articles, books, etc. These RS are based on the annotations (tags) of items to be recommended on the one hand and on the declaration or inference of user profiles on the other hand. In the tourism domain, this type of approach requires adaptations to integrate the spatiotemporal dimensions of POIs (Point Of Interest) in the process of recommending places to visit. Moreover, most of the existing works in the literature only use similarities between POIs to make suggestions to tourists and don’t calculate the prediction of future ratings given by tourists for the recommended POIs not yet visited. To overcome these problems, we propose three new content-based recommendation approaches for the tourism domain, the first one makes non-personalized recommendations using POI/POI similarity and the other two approaches make personalized recommendations based on inference of tourist profiles. In our work, we describe tourist profiles and POIs profiles using a list of tags defined by tourism domain experts. Then, the browsing history (ratings, comments) of mobile tourists is collected progressively from our website of tourist places of Chlef department for providing an online evaluation of our three approaches. The results of our experimentation allow us to compare our three approaches using the acceptance rate of POI suggestions by real users and show which approach mitigates the most the problems related to cold start and data sparsity.