Challenges in feature importance interpretation: Analyzing LSTM-NN predictions in battery material flotation

Abstract

Gomez-Flores et al. proposed a Long Short-Term Memory Neural Network (LSTM-NN) for predicting the flotation behavior of battery active materials using various physicochemical and hydrodynamic variables. While they achieved high prediction accuracy, validated through Mean Relative Error (MRE) and Mean Squared Error (MSE) metrics, concerns arise regarding the integrity of feature importance assessments derived from SAGE and SHAP methodologies. Specifically, the reliance on these model-specific techniques can introduce biases, obscuring the true relationships between features. Additionally, while Spearman's correlation elucidated significant relationships among variables, the absence of discussion on p-values left gaps in interpretation. This study emphasizes the need for cautious interpretation of feature importance metrics and the elimination of less significant variables, aiming to enhance model robustness and improve actionable insights in machine learning contexts.