Ensembles of Choice-based models for Recommender Systems

TítuloEnsembles of Choice-based models for Recommender Systems
Autor/aAmeed Ali Almomani
DirectoresEduardo Sánchez Vila
TipoTese doutoral
Data de lectura09/09/2020
Lugar de lecturaUniversidade de Santiago de Compostela
AbstractIn this thesis, we focused on three main paradigms: Recommender Systems, DecisionMaking, and Ensembles. For this purpose, this work is structured as follows.First, the thesis analyzes the potential of choice-based models. The motivation behindthis was based on the idea of applying sound decision-making paradigms, such as choice andutility theory, in the field of Recommender Systems. From this perspective, the recommenda-tion problem is considered a problem of choice prediction rather than rating prediction. Themotivation behind this work was to overcome some of the limitations of current state-of-artrecommender algorithms by providing: (1) accurate preferences, which are learnt from userchoices rather than from ratings, and (2) transparency, which is achieved by means of the setof estimated coefficients of the choice models.Second, this research analyzes the cognitive process underlying choice behavior. On theone hand, neural and gaze activity were recorded experimentally from different subjects per-forming a choice task in a Web Interface. On the other hand, cognitive were fitted usingrational, emotional, and attentional features. The model’s predictions were evaluated in termsof their accuracy and rankings were made for each user. The results show that: (1) the atten-tional models are the best in terms of its average performance across all users, (2) each subjectshows a different best model, and (3) ensembles may perform better than single choice modelsbut an optimal building method has to be found.Finally, the work explores the hybridization of choice-based models with ensembles. Thegoal is to take the best of the two worlds: transparency and performance. Two main methodswere analyzed to build optimal choice-based ensembles: uninformed and informed. For thefirst one, two strategies were evaluated: 1-Learner and N-Learners ensembles. The resultspoint out the superior performance of N-Learners ensembles and show the potential of per-sonalized arrangements. For the second one, we relied on three types of prior information: (1)High diversity, (2) Low error prediction (MSE), (3) and Low crowd error. A Greedy approachwas used to select the single models and to construct the ensembles. Again, two strategies,common and personalized models were analyzed in terms of their accuracy as well as the frequency of the best model. The results show the superior performance of informed ensembles,and indicate the low MSE as the key prior to build optimal choice-based ensembles.