Enhancing Corporate Tax Compliance and Fraud Detection Using Principal Component Analysis and Auto-Encoder

In recent years, corporate tax compliance has remained a contentious issue, particularly to the organ of record in the effort to oversee the compliance surge to ensure that fraudulent have not infiltrated the system. Currently, tax declarations are a form of self-assessment; hence, the marginal figures required proper mechanisms for identifying tax evasion. The proposed work in this study revolves on improving tax conformity and fraud identification by implementing Principal Component Analysis (PCA) and Auto-Encoder methods. This research is unique because of the use of two distinct approaches where the PCA is used for dimensionality reduction, and the AutoEncoder for anomaly detection which when used collectively facilitates more effective and efficient identification of fraudulent activities in tax data. Thus, a proposed framework was created: data preprocessing; feature selection with implementing Principal Component Analysis (PCA); and anomaly detection using AutoEncoder, which was trained on a dataset of 1,367 tax declarations for the Urban Delineation tax in Bogotá, Colombia. The outcomes revealed that the proposed Principal Component Analysis and Auto-Encoder (PCA-AE) model was accurate to 98.5% in the classification and had precision and recall ratios of 97&95% respectively which was consequently better than conventional machine learning algorithms like logistic regression and random forest. The results presented here suggest that the proposed framework may claim more negative incidents than formerly and boost tax revenue safeguard. Therefore, the study proves the feasibility of using hybrid PCA Auto-Encoder models, enabling enhanced corporate tax compliance and fraud identification, and further work on the development of automated tax audit methodologies.