Privacy-Aware Income Prediction Using Deep Neural Networks on the UCI Adult Dataset

Abstract

Income prediction is essential for applications in financial planning, credit scoring, and socioeconomic policy. In this study, we evaluate the effectiveness of deep learning neural networks—specifically Feedforward Neural Networks (FFN), Recurrent Neural Networks (RNN), and TabNet—for predicting income levels using the Adult dataset. We benchmark these models against traditional machine learning algorithms and investigate the impact of various preprocessing techniques, including missing value imputation, label encoding, categorical grouping, and normalization. Furthermore, we assess multiple feature selection methods—such as Chi-square, ANOVA, Random Forest importance, and L1 regularization—to determine their influence on predictive performance. Our experiments show that FFN achieves the highest baseline accuracy of 85.14% using optimized preprocessing, and further improves to 85.76\% when combined with feature selection techniques like Chi-square and ANOVA. These results confirm the advantage of deep learning approaches over classical models and underscore the value of well-designed data preprocessing pipelines. This study provides a comprehensive comparative analysis and highlights the importance of combining preprocessing with feature selection to optimize model accuracy in tabular data contexts. Future work will explore privacy-preserving training methods to enhance data protection when working with sensitive personal information.