Comparison of Feature Transformation and Missing Value Imputation Methods for Nutrient-Based Food Group Classification

Accurate classification of food items based on nutrient profiles is crucial for dietary analysis, nutritional research, and public health. This study investigates the effectiveness of various imputation and feature transformation methods in improving the performance of nutrient-based food group classification, particularly in the presence of skewed data and missing values. The dataset, sourced from the Tabel Komposisi Pangan Indonesia (TKPI) 2017, comprises 1,145 food items across 12 food groups, characterized by 20 nutrient features. We compared mean, median, K-nearest neighbors (KNN), iterative imputation, and iterative imputation with RandomForestRegressor for handling missing values, combined with feature transformations including standard scaling, log transformation, Box-Cox, and Yeo-Johnson. LightGBM with SMOTE was employed as the classification model, evaluated using macro-average accuracy to ensure equal importance to all classes. Our results show that median imputation paired with either Yeo-Johnson or Box-Cox transformations achieved the highest macro-average accuracy (0.8009), improving upon the baseline model (mean imputation with no transformation, accuracy 0.7721). The study revealed challenges in classifying certain food groups, particularly spices, highlighting limitations of nutrient-based classification for some categories. These findings suggest that appropriate imputation and transformation methods can improve classification accuracy in nutrient-based food group classification, while also indicating areas for future research to address persistent challenges.