DQM: an open source library to revolutionize data quality evaluation for industrial AI

Data quality is essential to ensure the reliability of AI models. The DQM library provides a concrete response to this challenge by offering relevant and interpretable quality attributes that assess critical aspects such as the representativeness and coverage of data within specific operational domains.

The institute’s teams developed two main categories of metrics:

1. Data-intrinsic metrics (e.g., representativeness, diversity), to evaluate dataset quality before an AI model is implemented;
2. System-dependent metrics (e.g., data-model coverage), which measure the impact of data on system performance once integrated into a model.

DQM was designed as a standalone Python package, making it easy to use independently or to integrate into other tools, such as DebiAI. The library has already been integrated into the end-to-end methodologies of major industrial players such as Naval Group and Valeo, strengthening their ability to accurately assess data quality.

Focus: concrete results with European and international reach

4 technology transfers to CAB project partners (RTE, Orange) and European players (Flatlandet, EnliteAI)

• InteractiveAI serves as the technological foundation for the H2020 AI4REALNET project, strengthening its position as a reference platform in the field of interactive AI
• The project has led to the publication of 8 scientific papers, contributing to the advancement of knowledge in AI and human-machine interaction