Abstract
Recommender Systems are software tools aimed at easing various types of decision-making processes. These technologies have been largely applied to the web. For instance, on an e-commerce platform, a target user may be suggested to purchases a selection of goods. Recommender Systems leverage the knowledge of user online behaviour, i.e., the user’s ratings or clicks. In this thesis, we investigate the possibility to generate suggestions to users while they move in the real physical environment: an aspect that has received much less attention so far. To reach our goal we leverage both users’ online and offline behaviour data, that is, coming from logging of user interactions both on web applications and with the physical environment. We operationalize the collection of offline user behaviour data by harnessing sensing-solution, such as, GPS and Internet of Things sensors, that allow collecting at a fine-grained level the user-environment interactions. Users’ online behaviour is considered when they upload pictures taken offline, as a tool for recording user interactions with the physical world. The main contribution of this thesis consists of a novel class of Recommender Systems that learn a generalized user behavioural model and generate recommendations on the base of that.
User behaviour learning is performed by grouping together users’ actions sequences in such a way that the resulting groups contain like-behaving/interested users. Then, by applying Inverse Reinforcement Learning a generalised user behaviour for each group of users is learnt. Recommendations of next-items to be consumed for each group of like-behaving/interested users are based on the group-specific learnt generalized user behaviour model (IRL-based model). We show that the learnt model is both structural, i.e., it explains the users’ sequential choices, and is generative, i.e., it can be used to generate new observations. We compare the proposed recommendation technique with state-of-the-art neural network-based, nearest neighbour-based recommendation algorithms and a content-based recommender that we also have designed with classical Machine Learning experiments and also in user-studies. We show that IRL-based models can offer to a user more interesting item recommendations: suggestions are both novel and relevant.