A Deep Learning Ordinal Classifier

Deep learning models such as TabNet have gained popularity for handling tabular data. However, most existing architectures treat categorical variables as nominal, ignoring the inherent ordering in ordinal data, which can lead to suboptimal classification performance, particularly in tasks where ordinal relationships carry meaningful information, such as quality assessment, disease severity staging, and risk prediction. This study investigates the impact of explicitly modeling ordinal relationships in deep learning by developing an ordinal classification model and comparing it with its nominal counterpart. The proposed approach integrates TabNet a deep learning framework with ordinal constraints, leveraging a proportional odds model to better capture the ordinal structure and Beta cross-entropy as the loss function to enforce ordering during training. To evaluate the effectiveness of the proposed ordinal classification approach, experiments were conducted on two publicly available datasets: the White Wine Quality dataset and the Hepatitis C dataset. The results demonstrate that incorporating ordinal constraints leads to improvements across multiple evaluation metrics, including 1-off accuracy, average mean absolute error (AMAE), maximum mean absolute error (MMAE), and quadratic weighted kappa (QWK) compared to a nominal classification model trained under the same conditions. These findings underscore the importance of ordinal modeling in tabular classification and contribute to the advancement of deep learning techniques for structured data.