Numerous scientific studies have established the high prevalence of psychiatric disorders worldwide. Despite the substantial medical needs they generate, diagnosis remains frequently delayed and prolonged due to the shortage of psychiatrists and the limitations of existing screening instruments. The identification of a psychiatric disorder generally relies on a self-report questionnaire, whose completion typically requires between five and twenty minutes. These constraints could be partially alleviated through the development of a digital screening tool that is comprehensive, reliable, efficient, and personalized.

Certain psychiatric disorders are associated with impairments in cognitive functioning. The design of a screening application therefore raises a broader question: how can one efficiently assess and synthesize the cognitive characteristics of an individual? This issue extends beyond the clinical setting. In the educational domain, the assessment of acquired competencies and the personalization of instructional support rely on similar challenges and data structures. Personalizing pedagogical assistance requires automatically estimating a student’s cognitive profile in order to provide guidance that adapts to their needs and to the knowledge inferred about them. Societal demand for personalized digital mentoring systems has grown substantially over the past decade, and the availability of large-scale data generated by educational platforms has made this area a particularly active field of research. In this work, we propose an algorithmic solution to the automatic and adaptive evaluation of individuals.

In both contexts, the central objective is to model an individual’s profile based on their responses to questions. Psychometrics and cognitive diagnosis have introduced mathematical models grounded in scientific knowledge from psychiatry, psychology, and pedagogy to address this problem. These approaches represent individuals in terms of latent cognitive traits. Although widely adopted, their predictive capacity remains limited. Recent advances in machine learning have significantly improved predictive performance. To achieve this, machine learning replaces part of the domain-specific assumptions and predefined structural constraints with more generic models whose parameters are optimized directly from data with respect to a predictive objective. However, the resulting individual profiles are often difficult for psychiatrists or educators to interpret and exploit in practice. Moreover, their reliability may be questioned, as it no longer relies explicitly on well-established theoretical assumptions.

The field of interpretability aims to restore semantic clarity and trust in individual profiles by expressing machine learning models in terms that are understandable to human experts. Recent work seeks to structure profiles into distinct and semantically coherent dimensions, such as hyperactivity or sleep disturbances. It also aims to ensure that the numerical values of a profile are meaningful to domain professionals, in particular by enforcing a monotonic consistency between observed responses and the corresponding latent scores. For example, a patient with a higher score on an “Impulsivity/Anger” dimension should indeed exhibit stronger manifestations of that tendency. Existing approaches enforce such properties either through constraints on model parameters or through specific data sampling strategies. However, these methods may reduce interpretability by increasing model complexity, without guaranteeing satisfactory monotonic behavior.

This thesis introduces a novel machine learning model based on a joint embedding of users and questions in a shared multidimensional space. We propose an original learning objective designed to explicitly enforce monotonicity constraints on user profiles with respect to their responses. These constraints are selectively applied to specific profile dimensions depending on the associated questions, thereby endowing the embedding space with structured semantic meaning. In addition, we introduce a second learning objective, inspired by Bayesian Personalized Ranking, to enhance predictive performance. We first address the case of binary responses, and subsequently extend the framework to ordinal and continuous responses. The generality of the proposed model enables its application to multi-target ordinal prediction, where it achieves state-of-the-art performance. To meet the requirements of adaptive testing, we further introduce a meta-learning algorithm that optimizes the iterative updating of an individual’s profile as new questions are administered.

The reliability of the proposed methods is established through theoretical analyses of the learning process. Their superiority in terms of both predictive accuracy and interpretability is demonstrated through quantitative and qualitative evaluations on real-world datasets from clinical and educational contexts. Finally, we demonstrate the feasibility of a concrete application of our algorithms to psychiatric disorder screening.

Publication(s) :
- Batel, A., Benouaret, I., Fruitet, J., Plantevit, M., & Robardet, C. (2024, October). A Simple Yet Effective Interpretable Bayesian Personalized Ranking for Cognitive Diagnosis. In ECAI 2024-27th European Conference on Artificial Intelligence, 19-24 October 2024, Santiago de Compostela, Spain-Including 13th Conference on Prestigious Applications of Intelligent Systems (Vol. 392, pp. 2386-2393). IOS Press. (paper )

-  A. Batel, C. Robardet, M. Plantevit, and I. Benouaret, ‘An Interpretable Model for Multi-Target Predictions with Ordinal Outputs’, Machine Learning, 2026, Accessed: Feb. 27, 2026. [Online]. Available: https://hal.science/hal-05525069

Code open source :

- CD-BPR algorithm : liris gitlab repository

- IMPACT algorithm: github repository