Multi Level Artificial Neural Tree Approach with Signature Database Integration for Textual and Image Data

Machine learning (ML) and deep learning (DL) applications are improving operational efficiency in several domains. Combining models from ML with DL provides not only a viable option but is also advantageous for real-world problems. In this paper, we propose an explainable artificial neural tree model that is a novel three-level tree architecture that comprises primary, intermediate, and target trees to process the textual datasets for classification and regression. Our methodology incorporates a neuron signature database that holds ample information on the intermediary tree structure of each neuron. We introduce a tree generation algorithm and a prediction algorithm that elaborate on the testing phase for textual data. To illustrate the efficacy of our model, we leverage two distinct datasets: a well-known publicly available iris dataset and a synthetic diabetes dataset. We continue to collaborate to build the tree structure for image processing. We present the image dataset's preprocessing and feature extraction procedures. We propose an algorithm for building trees for an image, which may be applied to image recognition by means of comparison with an example image. Moreover, we provided a prediction algorithm for the image recognition dataset. We practically demonstrated our model by generating explainable artificial neural trees and conducting the training and testing processes on two sample grayscale images of handwritten digits, 0 and 1. The trees for the training images contain 118 and 43 neurons, corresponding to the digits 0 and 1, respectively. For testing, a tree with 33 neurons was generated for an image representing the digit 1. Our proposed tree model leverages the accuracy of textual and image data, and the results of our experiments validate the efficacy of our model.