Data Transformation and Scaling Methods for Zero-Inflated Sparse Data Analysis

Zero-inflated sparse count data are common in text analytics, where keyword frequencies contain many zeros and highly skewed positive values. This study proposes a systematic model-specific framework that treats predictor transformation and scaling as explicit modeling factors and evaluates their effects across statistically different count and machine-learning models. Raw, log, square root, and binary transformations are combined with no scaling, standard scaling, and maximum-absolute scaling. The resulting pipelines are evaluated using ridge regression, random forest, negative binomial regression, zero-inflated negative binomial regression, and a two-part negative binomial model on extended reality patent keyword data with varying levels of sparsity. The results show that preprocessing effects differ substantially across model classes. Random Forest showed strong predictive performance under the full predictor setting and required little preprocessing, whereas continuous transformations provided modest improvements for Ridge and two-part NB models. Binary transformation with standard scaling substantially improves the predictive stability of negative binomial and zero-inflated negative binomial models by reducing extreme errors and convergence problems. These findings demonstrate that transformation and scaling should be selected jointly with the prediction model, response sparsity, and analytical objective rather than applied as universal preprocessing steps.