A novel generalized feature importance method

[Devashish13]

Hi Team,

I am pleased to introduce a generalized feature importance method that I have developed, inspired by the approach implemented in the h2o library. The h2o library's feature importance method relies on the weight matrix from the first two hidden layers of a trained neural network. In my work, I have devised a Python-based generalized version of this method, employing principles from matrix algebra to estimate feature importance. The mathematical formulation, which uses only one hidden layer, is detailed in this publication (https://www.worldscientific.com/doi/abs/10.1142/S0129065797000227).

To provide further insight into this novel approach, we have shared a preprint outlining the mathematical foundation of the proposed generalized feature importance method. You can access the preprint here (https://www.biorxiv.org/content/10.1101/2023.06.26.546510v1).

I have also implemented a Python function that can be used to estimate feature importance when provided with a trained neural network. The method employed in this function demonstrates remarkable speed compared to other existing techniques. Additionally, it has been rigorously validated on simulated data, successfully capturing true associations.

I am excited about the possibility of contributing this function to the Keras feature importance estimation methods section. Please guide me on the next steps in this regard.

Best Regards, Devashish Tripathi PhD Biostatistics National Institute of Biomedical Genomics India