🤖 machine learning

Explainable Artificial Intelligence for Anomaly Detection in Banking Transactions: An Internal Audit Perspective

This paper presents an Explainable AI framework combining Isolation Forest anomaly detection with SHAP-based explanations and a Streamlit dashboard to enhance transparency, precision, and auditor confidence in detecting fraudulent banking transactions.

Original authors: Anupa Lodhi

Published 2026-07-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Anupa Lodhi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the digital world of money as a massive, bustling city where billions of coins zip through invisible tunnels every single day. In this city, banks are the gatekeepers, trying to spot the pickpockets and thieves before they steal. For a long time, these gatekeepers used simple rulebooks: "If a transaction happens at 3 AM, stop it!" or "If the amount is over $10,000, stop it!" But the city is too big and the thieves are too clever; these old rules catch too many innocent people (false alarms) and, worse, they can't explain why they stopped someone. It's like a security guard shouting "Stop!" without saying what you did wrong.

This is where a new kind of science called Explainable Artificial Intelligence (XAI) comes in. Think of XAI not as a magic black box that just gives answers, but as a detective who not only solves the crime but also writes a clear, step-by-step report explaining exactly how they found the culprit. The paper you are about to read dives into this world, combining a smart pattern-spotting tool called Isolation Forest (which finds weird things by seeing how easily they can be separated from the crowd) with a mathematical method called SHAP (which acts like a magnifying glass, showing exactly which clues—like a strange time or a huge amount—made the computer suspicious). The goal is to help bank auditors, who are the human investigators, trust the computer's hunches and catch the bad guys without getting lost in confusing code.

The Paper: A Detective's Toolkit for Bank Auditors

In the paper "Explainable Artificial Intelligence for Anomaly Detection in Banking Transactions," author Anupa Lodhi tackles a tricky problem in the banking world. Banks are drowning in data, with billions of transactions happening daily, and they need to find the few that are fraudulent. The problem is that the smart computer programs (Machine Learning) that are good at finding these bad transactions are often "black boxes." They say, "This transaction is bad," but they can't tell the human auditor why. Without a reason, auditors can't write their reports, and regulators can't be sure the bank is doing its job.

The Solution: A Team of Three
Lodhi proposes a new system that acts like a three-person detective team, designed specifically for the internal audit teams at banks:

The Spotter (Isolation Forest): This is the first line of defense. Imagine a giant forest where every tree is a path. Normal transactions are like a crowd of people walking together; they take a long, winding path to get through the forest. Anomalies (the fraudsters) are like outliers; they are so weird that they get isolated quickly, taking a very short path. The Isolation Forest algorithm uses this idea to flag suspicious transactions without needing to know what "fraud" looks like beforehand. It just knows what "normal" looks like and spots the things that don't fit.
The Explainer (SHAP): Once the Spotter flags a transaction, the Explainer steps in. Using a method called SHAP (which is based on a branch of math called game theory), it breaks down the decision. It asks, "How much did the amount of money contribute to this flag? How much did the time of day contribute?" It gives a score for each clue, turning a scary computer number into a plain English sentence like, "This transaction is suspicious because the amount was huge and it happened at an unusual hour."
The Interface (Streamlit Dashboard): Finally, the team built a simple, colorful dashboard. This is the tool the human auditor actually uses. It doesn't require them to know how to code or understand complex math. They just upload their data, and the dashboard shows them a list of risky transactions, color-coded as "High Risk" or "Low Risk," with the SHAP explanations right next to them.

What They Found
The authors tested this system using a synthetic dataset (a fake set of 10,000 bank transactions created on a computer to mimic real life). They didn't use real bank data for this test, but they carefully injected 500 "fake" fraudulent transactions (5% of the total) to see if the system could find them.

Here are the results, straight from the paper:

The new system (Isolation Forest + SHAP) caught the fake frauds with 0.91 precision and 0.88 recall.
- Precision means that when the system said "This is fraud," it was right 91% of the time.
- Recall means it managed to find 88% of all the fake frauds that were hiding in the data.
When they compared this to other methods (like Autoencoders and One-Class SVMs), the new system performed better. For instance, the Autoencoder only reached an F1-Score (a mix of precision and recall) of 0.83, and the One-Class SVM was even lower at 0.77.
Crucially, adding the "Explainable" part (SHAP) did not slow down the system or make it less accurate. The "Standard Isolation Forest" without the explanation layer had the exact same detection scores (0.91 precision, 0.88 recall), proving that the explanations are a free bonus that doesn't hurt performance.

The Human Touch
The paper didn't just look at numbers; it asked real people if this actually helps. Five professionals with experience in banking audits looked at the explanations the system generated. They rated the quality of the explanations 4.3 out of 5.0. The auditors said that having these feature-level explanations gave them the "justificatory language" they needed to write their reports. Before this, they had to guess why a transaction was flagged; now, the computer tells them exactly which clues to write down.

Why This Matters (and What It Doesn't Do)
This paper suggests that we can make AI useful for bank auditors by making it transparent. It argues against the idea that we have to choose between a smart computer and a human who understands the decision. Instead, you can have both.

However, there are some limits to how sure we can be about these results. The study used a synthetic dataset, meaning the data was made up by the authors, not taken from a real bank's daily operations. While the results look promising, the paper admits that real-world banking data might be messier and more complex than this clean, fake version. The system was also tested with only four specific features (amount, time, frequency, and location change), whereas real banks might need to track hundreds of different things.

Despite these limits, the paper concludes that this framework is a practical step forward. It shows that it is possible to build a tool that is accurate, easy to use, and gives the human auditor the confidence to act. It's a bridge between the complex world of machine learning and the practical, rule-heavy world of banking audits, ensuring that when the computer raises an alarm, the human in charge knows exactly why.

Technical Summary: Explainable AI for Anomaly Detection in Banking Transactions

Problem Statement
The banking sector's rapid digital transformation, particularly in India, has led to transaction volumes reaching billions daily, creating a significant challenge for internal audit functions. While machine learning (ML) offers superior pattern recognition for fraud detection compared to traditional rule-based systems, its adoption in regulated environments is hindered by the "interpretability gap." Conventional models often function as black boxes, providing anomaly scores without the rationale required for auditors to document findings, act on alerts, or satisfy regulatory bodies like the Reserve Bank of India (RBI) and the European Union's GDPR. Current solutions struggle with high false-positive rates and lack the evidence-grade explanations necessary for compliance reviews.

Methodology
The paper proposes an end-to-end Explainable Artificial Intelligence (XAI) pipeline designed specifically for internal audit workflows. The system architecture comprises four sequential components:

Data Ingestion: Accepts CSV transaction files containing four normalized features: transaction amount, hour of transaction, 24-hour transaction frequency, and a binary location-change flag.
Anomaly Detection Engine: Utilizes an Isolation Forest (iForest) algorithm for unsupervised anomaly scoring. The model is trained without class labels, leveraging the principle that anomalies are structurally easier to isolate (requiring fewer recursive splits) than normal data. The anomaly score $s(x, n)$ is calculated based on the mean path length $E[h(x)]$ required to isolate a transaction.
Explainability Layer: Applies SHAP (SHapley Additive exPlanations) using the TreeExplainer. This layer decomposes the model's output for each flagged transaction into an additive sum of feature contributions ( $\phi_i$ ), grounded in cooperative game theory. This quantifies how much each feature shifts the prediction from the baseline, providing a mathematically consistent justification for the anomaly flag.
Audit Dashboard: A lightweight Streamlit interface that renders risk-tiered results (High/Low Risk) and feature attributions in a format accessible to non-technical auditors. The system operates locally to comply with data localization directives (e.g., RBI data storage rules).

Key Contributions
The paper advances the literature and practice in four specific ways:

Domain-Specific Integration: It unifies unsupervised anomaly detection (Isolation Forest) with post-hoc explainability (SHAP) into a single pipeline, rather than treating them as separate concerns.
Risk-Tiered Scoring: It introduces a calibrated scoring scheme (High Risk if score > 0.70, Low Risk otherwise) to support audit prioritization.
Auditor-Centric Interface: It delivers XAI capabilities through a no-code dashboard, removing the requirement for machine learning expertise at inference time and formatting outputs for direct inclusion in audit reports.
Empirical Benchmarking: It validates the approach against three unsupervised baselines (Autoencoder, One-Class SVM, and Standard Isolation Forest) on a synthetic dataset.

Experimental Results
The framework was evaluated on a synthetic dataset of 10,000 transactions (5% injected anomalies).

Performance: The proposed iForest-SHAP framework achieved a Precision of 0.91 and a Recall of 0.88 (F1-Score: 0.89). This outperformed the Autoencoder (F1: 0.83) and One-Class SVM (F1: 0.77).
Accuracy Preservation: The addition of the SHAP layer did not degrade detection performance; the Standard Isolation Forest (without XAI) achieved identical precision and recall scores.
Feature Importance: SHAP global analysis identified Transaction Amount as the primary driver of anomaly flags (Mean |SHAP| = 0.412), followed by Transaction Frequency (0.287).
Qualitative Feedback: A usability assessment with five banking audit professionals yielded a mean rating of 4.3/5.0. Participants confirmed that feature-level attributions provided the necessary justificatory language to draft audit findings, a capability absent in systems returning only scores.

Significance and Claims
The paper claims that this framework demonstrates the practical feasibility of operationalizing accountable, transparent AI within regulated financial environments, specifically addressing the complex regulatory landscape of the Indian banking sector. By coupling high-accuracy detection with human-interpretable justifications, the system bridges the gap between advanced ML capabilities and the operational constraints of internal audit teams (limited bandwidth, lack of data science expertise, and strict compliance mandates). The authors assert that this approach enables faster triage and more defensible documentation, aligning with operational risk management principles.

Limitations
The authors maintain a modest stance regarding the study's scope. They acknowledge that the use of a synthetic dataset may not capture the full distributional nuances of real-world Indian banking transactions. Additionally, the feature set is minimal; production deployments would likely require richer behavioral data (e.g., beneficiary network graphs). Future work is explicitly directed toward obtaining anonymized real-world data through banking partnerships.

The Paper: A Detective's Toolkit for Bank Auditors

Technical Summary: Explainable AI for Anomaly Detection in Banking Transactions

More like this