# The SHAP Explainer Model for Binary Classifiers Detecting Functional Groups in Molecules Based on FTIR Spectra

## TL;DR

This paper explains how a deep learning model detects chemical groups in molecules using FTIR spectra, making its decisions more transparent.

## Contribution

The study introduces SHAP values to interpret a CNN-KAN model's decisions in detecting functional groups from FTIR spectra.

## Key findings

- The CNN-KAN model's decisions are based on spectral regions known to impact functional group detection.
- SHAP values effectively trace which parts of FTIR spectra influence the model's binary classification decisions.

## Abstract

One of the disadvantages of deep learning models is the difficulty in understanding the premises on the basis of which the models make decisions. This might hinder the applicability of the model due to legal issues. In this paper, we investigate the decision process of CNN-KAN model trained to recognize chemical groups based on recorded FTIR spectra. The CNN-KAN model was trained as a binary classifier and using SHAP values, we could trace back the decision-making process and point out which part of the FTIR spectra are responsible for each positive or negative decision. It appears that the decision-making process of the deep learned CNN-KAN model is based on spectral regions that according to the literature have a large impact on the detection of particular functional groups.