Interpretable Feature Clustering for Dimensionality Reduction. An application to Explainable Remaining Useful Life Prediction

In prediction, variable selection allows to handle high-dimensionality, variable
redundancy, multicollinearity and instability. Classical approaches, such as
penalised supervised methods, like Lasso, may become unstable in the presence
of strongly correlated variables and rely on specific modeling assumptions.
Unsupervised feature clustering provides an alternative strategy for dimensionality
reduction that is independent of the response variable. Unlike Principal
Component Analysis (PCA) [1], which produces latent synthetic components
that are often difficult to interpret, clustering of variables aims to identify
homogeneous groups of correlated features and to represent each group by a
meaningful representative variable [5]. This approach improves interpretability
while preserving structural information in the data.
We consider the Interpretable Divisive Feature Clustering (IDFC) framework,
initially introduced for explainable Remaining Useful Life (RUL) prediction
[2]. IDFC combines divisive hierarchical splitting with clustering around latent
components [5, 4] in order to detect unidimensional groups of correlated
variables and select representative original features rather than synthetic components.
In this work, we extend this framework to more complex data, such as CMAPSS
turbofan engine datasets [3], characterized by higher dimensionality,
multiple operating regimes, and interacting fault modes. The proposed approach
enables robust dimensionality reduction while preserving semantic interpretability.
Experimental results show that IDFC, combined with a one-layer
LSTM model, achieves competitive predictive performance for RUL prediction
while significantly improving the stability and coherence of SHAP-based explanations.