BiRank vs PageRank: Using SNA on Company Register Data for Fiscal Risk Prediction

TL;DR

This study compares PageRank and BiRank algorithms on company network data to predict fiscal risks, demonstrating BiRank's superior accuracy and efficiency in identifying potential fraud cases for tax audits.

Contribution

The paper introduces a novel application of BiRank for fiscal risk prediction, showing its advantages over PageRank in network-based fraud detection.

Findings

BiRank outperforms PageRank in prediction quality and runtime.

BiRank achieves 16.38% precision on top 20,000 companies.

BiRank effectively identifies 83.4% of all fraud cases.

Abstract

Efficient financial administrations need to ensure compliant behavior of all tax subjects without excessive personnel costs or obstruction of compliant companies. To do so, accurate prediction of non-compliance or fraud is crucial. Social Network Analysis (SNA) provides powerful tools for fraud prediction as fraudulence is often clustered in certain areas of real world social networks. In this paper we present our results of comparing PageRank and the more recent BiRank to infer risk-ranks based on network structure and prior fraud information. Specifically, we model our social network from company register data. We find that in this case study BiRank outperforms PageRank in both quality of the resulting ranks for fraud prediction and run time. The results show that this class of algorithms is generally useful for fraud and risk prediction and more specifically also illustrate the potential of BiRank in comparison, as it opens up new modeling opportunities. Our results show that selecting companies for tax audits based on BiRank yields a precision of 16.38% for the top 20.000 subjects selecting 83.4% of all fraud cases (recall).