Title: Partial dependence and functional principal component-based reconstruction for explainable functional random forests Authors:  Fabrizio Maturo - Universita Telem.universitas Mercatorum (Italy) [presenting]
Annamaria Porreca - University San Raffaele Roma, Rome, Italy. (Italy)
Abstract: Functional random forests (FRF) combine the strengths of ensemble learning with functional data analysis, offering strong predictive performance on high-dimensional functional datasets. However, their limited transparency poses a major barrier in critical applications. This contribution introduces a set of explainability tools that extend classical partial dependence plots to the functional context, through functional partial dependence plots (FPDPs), enabling the study of the marginal effect of each functional principal component (FPC) score. To support interpretation, FPDPs are paired with graphical reconstructions of the functional shape associated with score variations. The approach highlights how individual FPCs influence predictions both in score space and in the time domain. Further, model-specific and model-agnostic FPC importance measures are provided, including an integrated visual tool comparing internal and external relevance. Applied to ECG signals, the method reveals meaningful patterns that support the interpretation of model behavior. These tools help bridge the gap between accuracy and explainability in functional machine learning.