Investigation of Learning Rate for Directed Acyclic Graph Network Performance on Dysgraphia Handwriting Classification
Siti Azura Ramlan, Iza Sazanita Isa, Muhammad Khusairi Osman, Ahmad Puad Ismail, Zainal Hisham Che Soh · 2023
Intelligence-based detection techniques are significant for dysgraphia diagnosis in school children at an early age. The method involves analyzing the symptoms through handwriting images. Directed Acyclic Graph (DAG) network is one of the intelligence-based technique in handwriting identification. The layer of DAG construction differs from the standard sequential convolution network that exhibits successful image classification. Recently, limited research has been conducted to investigate the learning rate that could affect the DAG network performance in classifying handwriting images. The issue of inconsistent handwriting patterns among dysgraphia children solved by the standard sequential convolution network layer remains different from the DAG operation. Therefore, this study aims to investigate the DAG operations with four acyclic blocks, along with varying the learning rates that could impact the performance of dysgraphia handwriting classification based on synthetic letter images. Experimentally, an improved DAG network model was investigated on four variant values of learning rate; 0.1, 0.01, 0.001, and 0.0001. The performance was measured using a confusion matrix for predicting dysgraphia or non-dysgraphia handwriting. The results obtained the best training accuracy of 99.01% produced by the DAG model at 0. 01learning rate. The testing performance for dysgraphia handwriting classification dropped at the lowest learning rate.