Publikation
Latent Boost: Enhancing Interpretability Through Loss-Defined Classification Objective in Structured Latent Spaces
Daniel Geißler; Bo Zhou; Mengxi Liu; Paul Lukowicz
In: Kohei Arai; Pascal Lorenz (Hrsg.). Intelligent Systems and Applications. Intelligent Systems Conference (IntelliSys), Amsterdam, Pages 91-114, ISBN 978-3-032-32726-0, Springer Nature Switzerland, 2027.
Zusammenfassung
Supervised machine learning often operates on the data-driven paradigm, wherein internal model parameters are autonomously optimized to converge predicted outputs with the ground truth, devoid of explicitly programmed rules or a priori assumptions. Although data-driven methods have yielded notable successes across benchmark datasets, they inherently treat models as opaque entities, limiting interpretability and explanatory insights into their decision-making processes. Moreover, existing distance metric learning methods are primarily designed for clustering or retrieval tasks and are not effectively integrated into end-to-end supervised classification. In this work, we introduce Latent Boost, a training methodology that explicitly incorporates latent space structure into the classification objective. We combine probabilistic cross-entropy loss with distance metric learning through a weighted loss formulation, enabling simultaneous optimization of prediction accuracy and latent cluster organization. Thus, the model is not only optimized for classification of discrete data points but also enforces compact and well-separated class representations. By leveraging structural insights from intermediate latent representations, Latent Boost improves interpretability, as demonstrated by higher Silhouette scores, while accelerating training convergence. These benefits are achieved with minimal additional cost, making Latent Boost broadly applicable across datasets and architectures without data-specific adjustments. By enabling models to self-organize their internal representations during training, Latent Boost acts as a structural regularization mechanism aligning classification performance with transparent latent structure, positioning it as a general framework for interpretable and efficient classification systems.