Abstract
Deep neural networks have shown, over the last decade, a significant increase in performance for many predictive tasks. Nevertheless, their complexity (large number of trainable parameters) make them appear as black boxes to the users, with the latter having difficulties in the understanding of their predictions. Recently, the interest of the scientific community in interpreting their predictions is significantly increased. In this thesis, we focus on data visualization techniques that leverage importantly the improvement of their interpretability.First, we propose a new visualization approach to explain the predictions of neural networks in a binary text classification setting, that is addressed to regular as well as to users who are neural networks experts. This visualization is based on the construction of localities in the text representation space. The visualization of the decision boundary is then made using distances similar to those present the representation space.We also propose a second approach for presenting the predictions of neural networks, for a multiclass text classification task. It is addressed to users who are not neural network experts. This module focuses primarily on a case study: the investigation of mentions, in the social media, of non-pharmaceutical interventions used that are linked to cancer.In general, this work has opened promising perspectives on the production of explanations based on the decision boundary and on the exploration of data from social media.