Combination of User and Venue Personality with Topic Modelling in Restaurant Recommender Systems

Recommender systems are popular information systems used to support decision makers' information overload. However, despite their success in simple problems, such as music recommendation, they have been criticized of insufficient performance in highly complex domains, characterized by many parameters, such as restaurant recommendations. Recent research has acknowledged the importance of personality in influencing consumers' choice, but recommendation methodologies do not exploit this in the restaurant recommendation problem. Hence, this work seeks to analyze the contribution of personality in combination with extracted topics from consumers' electronic word of mouth (eWOM) to restaurant recommender systems. The paper utilizes a bi-directional transformer approach with a feed-forward classification layer for personality prediction, due to its improved performance in similar problems over other machine learning models. One issue with this approach is the handling of long text, such as narratives written by people of different personality types (labels). Thus, different long-text management methods are evaluated to find the one with best personality prediction performance. Two personality models are evaluated, namely the Myers-Briggs and Big Five, based on two labelled datasets that are utilized to generate two personality classifiers. In addition to customer personality, this work investigates the concept of venue personality estimated from personalities of users that visited a venue and liked it. Finally, the customer and venue personalities are used together with the topics discussed by customers to form the input to the extreme gradient boosting (XGBoost) models for predicting user ratings of restaurants. The performance of these models is compared to traditional collaborative filtering methods using various prediction metrics.