Artefact Research Center

A widely used family of machine learning models is based on decision trees: random forests, boosting. While their accuracy is often state of the art, such models suffer from a black-box feeling, giving limited control to the user. We aim to increase their explainability and transparency, typically by improving the estimation of SHAP values in the case of unbalanced datasets. We also aim to provide some guarantees for such models, e.g., for out-of-training samples or by enabling better monotonic constraints.

The challenge of better adoption of AI in companies is to improve the AI models on the one hand, and to understand the human and organizational aspects on the other. At the crossroads of qualitative management research and social research, this axis seeks to explore where businesses face difficulties when adopting AI tools. The existing frameworks on innovation adoption are not entirely suitable for machine learning innovations, as there are typical differences with regulation, people training or biases when it comes to AI, and more so with generative AI.

The project will mobilize qualitative and quantitative research methods and address two key questions: How can companies effectively measure social and environmental sustainability performance? Why do sustainability measures often fail to bring about significant changes in organizational practices?

On the one hand, the project aims to explore data-driven metrics and identify indicators to align organizational procedures with social and environmental sustainability objectives. On the other hand, the project will focus on transforming these sustainability measures into concrete actions within companies.

When a model makes a prediction based on an image, for instance showing a face, it has access to sensitive information, such as the ethnicity, gender or age, that can bias its reasoning. We aim at developing a framework to mathematically measure such bias, and propose methodologies to reduce this bias during the model training. Furthermore, our approach would statistically detect zones of strong bias to explain and understand and control where such models reinforce the bias present in the data.

One major application of LLMs is when coupled with a corpus of documents, which represent some industrial knowledge or information. In such a case, there is a step of information retrieval, for which LLMs show some limitations, such as the size of the input text, which is too small for indexing documents. Similarly, the hallucination effect can also happen in the final answer, which we aim at detecting using the retrieved document and model uncertainty at inference time.

Besides our team dedicated to research, we have several collaborators who spend some time doing scientific research and publishing papers. By working also as consultants inspire them with real-world problems encountered with our clients.