Determining Optimal Granularity for Effective Handwriting Analysis

In most literature on handwriting analysis, the datasets used are typically mentioned. However, the sizes of these datasets can vary once they undergo preprocessing for actual model training. This results in different sample sizes being used as inputs. We hypothesize that these varying input sizes can influence the output results of the models. In this paper, we explore the optimal granularity for handwriting analysis. We trained two deep learning models to classify traits such as extraversion (EXT) and conscientiousness (CON) using our own dataset. Our findings indicate that the optimal granularities are 3 × 8, 8 × 6, and 9 × 12 for different splitting patterns. We recommend selecting training samples with at least 77 instances, each containing 3 to 4 lines of text, to ensure robust model performance. These guidelines can serve as a reference for future research in handwriting analysis.