A New Risk-Sensitive Deep Learning Optimization Function for Ranking Tasks

This master thesis proposes the RiskLoss function to deal with the (hard) problem of incorporating risk-sensitiveness measures into Deep Neural Networks (DNNs), by including two adaptations for neural network ranking in ad-hoc retrieval and Recommender Systems (RSs): a differentiable loss function and the use of networks‘ sub-portions, obtained via dropout, as baseline systems for optimizing risk sensitiveness. We empirically demonstrate significant achievements of the RiskLoss functions when used with recent DNN methods. For ad-hoc retrieval, RiskLoss presents the most consistent risk sensitiveness behavior, reducing by 28% the losses over the best evaluated baselines and significantly improving over the risk-sensitive state-of-the-art non-DNN method (by up to 13%) while keeping (or even increasing) overall effectiveness. In RSs, RiskLoss reduced the number of bad recommendations by over 11% for ”hard-to-recommend”users. This dissertation produced a full paper in the most important worldwide Information Retrieval (IR) conference (ACM SIGIR Conference on Research and Development in IR 2022, Qualis A1) and a journal paper submitted to Elsevier´s Information Processing & Management (Qualis A1), currently in the second round of revisions