Interactive Machine Learning Applied to Weighted Random Forests

Défense de mémoire de Romain Lahaye

This thesis explores an approach still little studied : making the Weighted Random Forest model interactive. Although current learning models are generally effective, they often suffer from a lack of transparency
and flexibility, which limits their use in sensitive contexts where human intervention is essential. Moreover, certain important human abilities, such as intuition or common sense, can play a key role in guiding or adjusting these models, as long as an adequate interaction framework allows them to express themselves effectively. The main objective of this work was to create an interface that allows a user, even a non-expert, to interact
with the model to influence its functioning and improve its performance. An analysis of previous works made it possible to identify several good practices to guide the design of this interface, emphasizing data visualization,
clarity of feedback, and ease of interaction.

The developed application offers several types of interaction with the weighted forest : removal of lowperforming trees, addition of specialized trees trained on misclassified data or built from under-represented
features, as well as weight adjustment to guide the creation of new trees. Early tests show that some of these interactions can have a measurable effect on the model’s performance and structure. This work constitutes a
first exploration of the possibilities for interaction with weighted random forests, and opens up perspectives for the development of more accessible and better-adapted interactive tools for this type of model.

Keywords : interactive machine learning, weighted random forest

Contact : Isabelle Daelman - isabelle.daelman@unamur.be
Télecharger : vCal