Evaluating the Impact of Removing Low-relevance Features in Non-retrained Neural Networks

Feature selection is a widely used technique to boost the efficiency of machine learning models, particularly when working with high-dimensional datasets. However, after reducing the feature space, we must retrain the model to measure the impact of the removed features. This can be inconvenient, especially when dealing with large datasets of thousands or millions of instances, as it leads to computationally expensive processes. To avoid the costly procedure of retraining, this study evaluates the impact of predicting using neural networks that have not been retrained after feature selection. We used two architectures that allow feature removal without affecting the architectural structure: FT-Transformers, which are capable of generating predictions even when certain features are excluded from the input, and Multi-layer Perceptrons, by pruning unused weights. These methods are compared against XGBoost, which requires retraining, on various tabular datasets. Our experiments demonstrate that the proposed approaches achieve competitive performance compared to retrained models, especially when the removal percentage is up to 20\%. Notably, the proposed methods exhibit significantly faster evaluation times, particularly on large datasets. These methods offer a promising solution for efficiently applying feature removals, providing a favorable trade-off between performance and computational costs.