A methodology for approximating motivation-related latent states in large scale scenarios: and its role in engagement prediction within a video game setting

Motivation is a fundamental psychological process guiding our everyday behaviour. For doing so, it heavily relies on the ability to attribute relevance to potentially rewarding objects and actions (i.e., incentives). However, despite its importance, quantifying the saliency that an individual might attribute to an object or an action is not an easy task, especially if done in naturalistic contexts. In this view, this thesis aims to outline a methodology for approximating the amount of attributed incentive salience in situations where large volumes of behavioural data are available but no experimental control is possible. Leveraging knowledge derived from theoretical and computational accounts of incentive salience attribution, we designed an Artificial Neural Network (ANN) tasked to infer a latent representation able to predict duration and intensity of future interactions between individuals and a series of video games. We found video games to be the ideal context for developing such methodology due to their reliance on reward mechanics and their ability to provide ecologically robust behavioural measures at scale. We developed and tested our methodology on a series of large-scale ($N> 10^6$) longitudinal datasets evaluating the ability of the generated latent representation to approximate some functional properties of attributed incentive salience. The present work opens with an overview of the concept of motivation and its interconnection with engagement in a video-game setting. It proceeds by formulating the theoretical and computation foundations on which our methodology is built upon. It then describes the iterative process of model building, evaluation and expansion underlying the implementation of our methodology. It continues by analysing the latent representation generated by the ANN and comparing its functional characteristics with those of attributed incentive salience. The manuscript ends with a general overview of the potential applications of our methodology with a particular focus on the area of automated engagement prediction and quantification in videogames settings.