Explainability has become a central concern in modern machine learning, particularly for unsupervised models whose internal representations are difficult to interpret. This paper introduces Feature Maximization (FMax) as a general-purpose explainability metric designed to identify the most discriminative and representative features associated with any learned partition of data. Unlike classical approaches such as TF-IDF or pointwise mutual information, FMax combines a recall-oriented measure of feature coverage with a precision-oriented measure of feature exclusivity, yielding a unified score that reliably highlights the features that best characterize each cluster or topic. The metric is model-agnostic and can be applied to any system that produces a discrete assignment of instances to topics, clusters, or predefined classes, making it equally relevant for unsupervised clustering, topic modeling, and supervised classification tasks. Crucially, FMax is not limited to textual data: it has also been shown to enhance the interpretability of graph-based representations such as those produced by SINr, and can serve as an effective substitute for standard distance measures when dealing with highly multidimensional and sparse data spaces where Euclidean or cosine metrics tend to lose their discriminative power. We demonstrate the versatility of FMax through two contrasting paradigms of topic modeling. In the statistical setting, FMax is applied to outputs from models such as LDA and NMF, where it provides sharper and more coherent topic descriptors than standard baseline metrics. In the LLM-based setting, we apply FMax to topic representations generated by large language models, where the metric serves both as an evaluation tool and as a means to surface the most informative tokens driving each topic assignment. Comparative experiments on benchmark corpora show that FMax consistently improves human interpretability scores and aligns well with expert judgment across both paradigms. Beyond topic modeling, the generality of the metric suggests broad applicability to any classification or clustering scenario where post-hoc feature-level explanation is needed. This work thus positions FMax as a practical and theoretically grounded bridge between unsupervised learning and explainable AI.

Prof. Jean-Charles Lamirel holds a permanent position at IUT Robert Schuman, University of Strasbourg, where he has taught Information Science, Library Science, and Computer Science since 1997. He holds a doctorate in Computer Science (1995) and a Research Accreditation in the same area (HDR, 2010) from the University of Nancy. He is also an Associate Researcher at the LORIA laboratory (INRIA, Nancy) and at the SAMM laboratory of Université Paris 1 Panthéon-Sorbonne since 2021.

His research lies at the crossroads of artificial intelligence (Section 27) and information and communication sciences (Section 71), a distinctive interdisciplinary positioning that underpins his entire scientific career. His core expertise is in unsupervised neural learning, with a sustained focus on Self-Organizing Map (SOM) architectures and the development of the NOMAD-MultiSOM system — one of the first document retrieval systems to implement a full systems approach based on neural clustering. His methodological contributions span multiview data analysis, incremental learning, topic modeling, feature maximization metrics for semantic coherence, novelty detection, and, more recently, Retrieval-Augmented Generation (RAG) systems and LLM-integrated knowledge extraction.

A unifying concern across all of his work is explainability: ensuring that AI models produce interpretable and actionable results, particularly in the context of information science applications. His most significant applied domain is scientometrics, where his methods have enabled sophisticated bibliometric analysis, research trend detection, and science mapping deployed across multiple institutional and international settings. He is the author of more than 180 international publications, including 25 papers in peer-reviewed international journals, and has presented 91 invited talks at international conferences and seminars.

Prof. Jean-Charles Lamirel has built an extensive international academic presence. Since 2016, he has held a Sea-Sky Visiting Professorship at Dalian University of Technology (DUT), China, where he teaches Artificial Intelligence at the WISELab laboratory. He is the only non-Chinese permanent member of CAASP (Speciality Committee of Science of Science and Discipline Construction, China). His international collaborations also extend to Taiwan (National Taiwan University), Poland (Silesian University of Technology), Estonia (University of Tartu), Australia (University of Technology of Sydney) and Canada (UQAM, Montréal), among others.

His service to the scientific community is extensive. He serves as Area Chair (meta-reviewer) at leading NLP and data mining conferences including EMNLP and ICDM, and sits on the editorial boards of major journals such as Neural Computing and Applications (NCAA) and Knowledge and Information Systems (KAIS). He has participated in over 107 international program committees and chaired numerous conference sessions. He has co-supervised 12 completed doctoral theses (with a cumulative co-supervision rate of 630%), three post-doctoral researchers, and 23 Master's and engineering internships. He is currently co-supervising three additional doctoral projects.

As a researcher and teacher, Prof. Jean-Charles Lamirel embodies a rare dual commitment: advancing the theoretical foundations of intelligent data analysis while rigorously applying these methods to real-world challenges in information science, digital libraries, and knowledge management. His work bridges disciplines, institutions, and continents — contributing to both the computational and humanistic dimensions of the information society.