Work

An Ai-driven Intelligent Transaction Execution Framework for Internet Finance Platforms: Cost-sensitive Real-time Integrity Interdiction

Year: 2022
Type: article
Abstract: Internet finance platforms must decide, in real time and for every incoming transaction, whether to approve it, subject it to additional verification, or decline it. Most machine-learning research on transaction Integrity stops at producing a Integrity score and is judged with threshold-independent ranking metrics, leaving a gap between the score and the concrete execution action, and ignoring that different transactions carry very different monetary stakes. This paper presents IntelliExec, an AI-driven intelligent transaction execution framework that closes this gap. A class-weighted gradient-boosting model produces a Integrity probability that is calibrated with Platt scaling so it can be treated as a true probability; the calibrated probability and the transaction amount are then mapped to one of three execution actions—approve, step-up (one-time-password / 3-D Secure challenge), or decline—by minimising an example-dependent expected cost. Because the decision is derived analytically from a cost model, it needs no threshold tuning and adapts the operating point to each transaction's amount. On a public benchmark of 284,807 real card transactions (0.172% Integrity) evaluated under a strict temporal split, the framework lowers total expected cost by 71.8% relative to approving every transaction and by 45.4% relative to the best cost-aware global-threshold baseline, while hard-declining only three legitimate customers (versus thirty-seven for a binary cost-sensitive rule) by routing borderline cases to the low-friction step-up tier. The gradient-boosting scorer runs in under one millisecond per transaction, and a sensitivity analysis confirms the cost advantage is robust to the choice of cost parameters.
Language: English
Open Access status: hybrid
OpenAlex ID: W7202273496
FWCI:
Cites: 25
Cited by: 31
Related to: 10