Implementing AI-Driven Decision Support Systems in Data Management

Some stratification occurs since the graph clearly indicates that most fraudulent transactions occur in the high-risk zone. This supports the claim that the model can generate meaningful confidence measures, unlike binary decisions, and will support triage-based processes. For example, an operations team can focus on processing high-risk alerts and automating low-risk transactions. Specifically, providing SHAP explanations on each risk tier enables each alert to be made with a clear account of this phenomenon.

This simulation demonstrates that the real-world application of the AI-driven DSS could be performed in real time, facilitating proactive fraud mitigation, load-balanced alert elimination, and data-driven decision-making pathways. Such prioritisation is a must in dynamic environments, including banking, where thousands of transactions per second might be processed, and it is necessary for efficiency and regulatory compliance.

3.6. Cost-Sensitive Threshold Selection

Fraud detection is characterised by asymmetric misclassification costs, where a false negative (missed fraud) incurs substantially higher loss than a false positive (unnecessary manual review). To account for this imbalance, expected loss was modelled as:

where and denote the number of false negatives and false positives, respectively, and and represent their associated costs.

Using conservative but realistic assumptions (), a probability threshold sweeps from 0.01 to 0.90 was conducted using the calibrated XGBoost model. The analysis identified 0.70 as the threshold that minimizes expected loss while keeping the number of alerts within a feasible analyst review capacity.

Accordingly, final operational thresholds were defined as follows and are used consistently throughout the DSS:

• High Risk: ≥ 0.70
• Medium Risk:30 – 0.69
• Low Risk: < 0.30

These thresholds replace all earlier illustrative values to ensure internal consistency across evaluation, simulation, and deployment-oriented analyses.

3.7. Ablation Study: SMOTE vs Class Weighting

To assess the robustness of the proposed imbalance-handling strategy, three alternative configurations were evaluated and compared against the baseline SMOTE-XGBoost model:

1. XGBoost with scale_pos_weight only
2. SMOTE–Tomek Links
3. SMOTE–ENN

Given the very small number of fraud cases in the test set (n₊ = 7), the evaluation focuses on PR–AUC as the primary metric, supplemented by recall at a low false-positive operating point. Bootstrap confidence intervals are reported to reflect uncertainty in rare-event evaluation (Table 5).

Table 5. Ablation study of imbalance-handling strategies (n₊ = 7).

The results indicate that SMOTE–ENN achieves a marginally higher PR–AUC, while scale_pos_weight yields slightly higher recall at a strict false-positive constraint. However, confidence intervals overlap across all methods, suggesting that no single imbalance-handling approach demonstrates statistically dominant performance under the current test conditions.

Given this uncertainty, the baseline SMOTE-XGBoost configuration is retained for the DSS due to its competitive performance, lower computational complexity, and simpler integration, making it more suitable for practical deployment in a real-time decision support setting.

4. DISCUSSION

This section is a critical discussion of the findings of the study in terms of placing them in the context of the larger body of literature on AI-based Decision Support Systems (DSS), financial fraud detection, and explainable machine learning. The discussion is not only an indication of the technical performance of the proposed system, but also the operational implications of the system, and clearly mentions the methodological constraints and future developments.

4.1. AI-Driven DSS Performance in Data Management

The findings prove the point that the use of machine learning models in the context of a DSS can help to enhance fraud prioritisation in the environment with excessive class imbalance and sophisticated data, as it has been previously observed in the field of AI-enhanced decision support studies [1, 3]. Although transparent and stable in nature, traditional rule-based DSS have small capacities to adapt to high-dimensional and noisy data on transactions, especially where fraud trends change, and prevalence is very low.

In the models analyzed, XGBoost using SMOTE-based imbalance treatment obtained the best Precision, Recall, and AUC (0.568) than both Random Forest (0.512) and Logistic Regression (0.341). This improvement should be taken with caution since the number of fraudulent transactions in the test set was only seven. Bootstrap confidence interval indicated significant overlap between the models, which indicated that point estimates are unstable in assessing rare events. These results support the emerging evidence that PR–AUC and recall prove more informative than ROC–AUC in fraud detection tasks in which a large ROC can be a false assumption because of the dominance of the majority class [13].

Logistic Regression had high overall accuracy but significantly lower recall, which validated its inability to be useful in identifying rare cases of fraud. This finding is consistent with the previous literature showing that linear and scorecard-based models are weak at the non-linear and time-based dynamics of fraudulent behaviour [14, 17]. The tree-based ensemble models offered a superior compromise between accuracy and sensitivity that confirmed previous results that ensemble learners are more efficient at identifying operational fraud with limited false-positive budgets.

Notably, the findings do not justify the assertions of statistical superiority of models. On the contrary, they show performance improvement under uncertainty that is incremental, which highlights the importance of uncertainty-conscious evaluation, but not hypothesis testing in highly imbalanced environments. Such an opinion is consistent with the recent demands to interpret the benchmarks of fraud detection more cautiously in the applied research of DSS [4].

4.2. Explainability as a Trust Enabler in DSS

The addition of SHAP-based explainability to the DSS overcame a major obstacle to the application of AI models in the regulated setting, where the decisions need to be transparent and auditable [15, 16]. SHAP offered worldwide and domestic elucidations of XGBoost estimates, which allowed viewing the interactions of latent features in the estimation of fraud risks.

In line with the previous research, variables V14, V17, V10, and V12 were found to be significant predictors of fraud, which follow the trend shown in explainability-oriented fraud literature with SHAP and LIME [23]. Visual explanations can be used to achieve the goals of governance and auditability, as they contribute to the claim to explainable AI as a transition between automated systems and humans making decisions [19].

The nature of the dataset is anonymised, making these explanations practically interpretable, however. Since the features generated by the PCA do not have direct semantic meaning, SHAP explanations offer technical indisclosure as opposed to practical behavioural information. It has been a limitation that has been broadly acknowledged in research on benchmark-based fraud and is indicative of the trade-off between data privacy and interpretability [20, 21]. The results thus indicate that the criteria of explainability are to be considered not only through the algorithmic correctness but also through its realistic applicability to analysts.

4.3. Explainability Improvements and PCA Regressions

Although SHAP is both locally and globally interpretable, it relies on input feature semantic clarity to be successful. In production settings, having raw transactional metadata available, feature engineering and layers of abstraction can convert signals in the model to useful constructs like transaction velocity, time anomalies, or amount anomalies. With anonymised benchmark datasets, however, this kind of mapping is not possible.

Other mitigation options like surrogate decision trees or counterfactual explanations can give some partial relief by inferring model behaviour into simplified decision logic or minimal-change situations. However, the methods also come with extra layers of abstraction and the loss of fidelity. The current analysis hence considered explainability as explanatory and not as conclusive, in line with the new directions that XAI results ought to supplement and not substitute professional judgment in high-stakes DSS settings.

4.4. Operationalization and Latency Factors

Controlled conditions were used to test the indicative real-time performance of the DSS. XGBoost median latency of inference took about 0.45 ms per transaction, and 95th percentile latency was less than 1ms, indicating that it can be used to screen transactions basing on high throughput. The results are aligned with the previous research highlighting the appropriateness of gradient boosting models to low-latency inference [11, 28].

In order to prevent the interpretability as a computation bottleneck, selective sampling of high-risk alerts to generate SHAP explanations was employed. It is an asynchronous approach that complies with industry practice of achieving a balance between transparency and operational efficiency. Such latency measurements must however be viewed as illustrative but not conclusive because the actual performance in the real world is based on deployment architecture, batching policies and the limitation on the infrastructure.

4.5. Risk-Based Triage and Decision Support Value

The risk-tiering simulation has shown how operationally meaningful risk categories can be derived as the so-called calibrated model outputs, and the analysts can prioritise the investigations. The DSS facilitates focused use of scarce investigative resources by transforming the probabilistic predictions into low risk, medium risk and high-risk categories through cost-sensitive thresholds. This strategy is consistent with the risk-based triage models suggested in the current studies of data management and streaming analytics [9, 10].

This stratification leads to less fatigue in alerts and higher confidence in AI results, which concerns the previous studies of financial frauds about over-alerting and analyst overload [29, 30]. Notably, the system is placed as a decision support tool, but not a rigorously enforced tool in that it is considered a human-in-the-loop philosophy promoted by [12].

4.6. Governance, Monitoring, and Responsible Deployment

To ensure successful implementation of AI-based DSS, there must be constant monitoring and management of the use of the system to reduce the decline in performance over time. Population stability indicators, divergence-based drift detection, and periodic recalibration are the techniques that should be used to ensure reliability, as discussed by Al [8, 13]. Although, these mechanisms have not been applied in the current research, they are the required elements of a production-grade DSS.

Anonymised dataset ensured privacy protection on the experimentation, which is in line with best practices in responsible AI research [31, 32]. Nevertheless, assertions on encryption, adversarial-invariance, or sophisticated privacy-preserving privacy-preserving methods are theoretical, and ought to be proven empirically in subsequent research. Such careful phrasing is an indication of the necessity to strike a balance between innovation and regulatory and ethical responsibility in financial AI systems.

In general, the results indicate that AI-based DSS can improve the prioritisation and decision support of fraud in case it is developed in terms of imbalance awareness, uncertainty quantification, explainability, and operational realism. Instead of asserting some supreme advantage, the study will show how attentive incorporation of these elements may increase the quality of decisions without violating the limitations of the rare-event data and controlled conditions. Such reflections lead to the current transformation of DSS as being rule-based to adaptive, transparent, and human-centred AI-assisted decision-making systems.

FUTURE DIRECTION

Despite the high performance of the proposed AI-driven DSS in simulation, several limitations should be considered. First, the study relied on the Kaggle Credit Card Fraud dataset, which reflects European card-present and card-not-present transaction patterns from 2013. Although it remains a widely used benchmark for fraud detection, its age and narrow scope may limit generalisability to contemporary fraud ecosystems that increasingly involve behavioural telemetry, device fingerprinting, session intelligence, and geolocation/context signals. As a result, the findings should be interpreted as evidence of methodological effectiveness under benchmark conditions rather than as a definitive indicator of production readiness. Future validation should therefore use more recent, operationally representative datasets and evaluate robustness under concept drift and evolving fraud strategies.

Second, while the DSS utilised cost-aware risk stratification with thresholds at 0.30 and 0.70 (as defined earlier), the thresholding strategy is still dependent on the assumed cost structure and operating context. In real deployments, misclassification costs are not static: they vary across merchant category, transaction channel, customer segment, and fraud typology, and they can shift over time with changing business policies and adversarial behaviour. Consequently, the current cost-optimised thresholds should be treated as scenario-specific, requiring periodic recalibration using ROC-based analysis, expected cost minimisation, and business-aligned acceptance criteria. Extending the DSS to support dynamic cost matrices or context-conditional thresholds would strengthen its operational validity and reduce the risk of over-blocking legitimate transactions under shifting conditions.

Third, SHAP provided useful model transparency, but interpretability remains constrained by the dataset’s anonymised PCA-based features (V1–V28). Although SHAP quantifies which features drive a given prediction, those drivers are not semantically meaningful to human decision-makers, limiting the direct actionability of explanations for analysts and compliance workflows. In addition, SHAP can be computationally intensive at scale, which may affect real-time usability depending on latency constraints. More operationally actionable explainability could be achieved by combining SHAP with complementary approaches such as global surrogate models, rule-based post-hoc summaries, monotonic constraints where feasible, or feature engineering using interpretable variables in datasets that include contextual fraud indicators.

CONCLUSION

This paper tested an AI-driven Decision Support System (DSS) in detecting credit card fraud in case of extreme imbalance in classes and the goal of the research was to test whether current machine learning models could be used to support traditional decision-support systems and remain transparent and feasible to run. Based on the publicly accessible Kaggle Credit Card Fraud dataset, the proposed architecture incorporated the elements of class imbalance reduction, uncertainty-sensitive classification, explainability, probability calibration, and cost-sensitive risk scoring in a single DSS pipeline.

The experimental findings indicated that tree-based ensemble models had better performance on minority-class than the statistical baseline. XGBoost has the highest Precision Recall AUC (0.568) as opposed to the Random Forest (0.512) and Logistic Regression (0.341) and had a similar recall (0.571) to the Random Forest on the imbalanced test set. Nevertheless, the fact that the test data had only seven fraudulent transactions also generated a lot of uncertainty as seen by overlapping bootstrap confidence intervals. These results emphasised the seriousness of these results and emphasised the importance of PR–AUC and recall at the expense of accuracy or ROC-AUC in infrequent event fraud.

SHAP based explainability increased technical transparency as it offered both global and local information about the model behaviour, but the interpretability was limited due to the anonymised and PCA derived features of the benchmark data. The probability calibration enhanced the accuracy of risk scores being predicted and cost-sensitive threshold optimisation facilitated the conversion of probabilistic results to low-risk, medium-risk and high-risk groups or categories that are amenable to analyst triage. This showed that AI models can be used to aid in decision-making by ranking high-risk cases without straining operational capacity.

Thus, the research revealed that AI-based DSS may support, and not supersede, conventional fraud detection systems through the provision of balanced risk measures, explainability, and uncertainty-sensitive evaluation. Even though the findings were constrained by the range of datasets and the instability of rare events, the framework offered a clear, reproducible, and operationally sound basis of applying AI-supported decision support in controlled financial settings. This method should be applied to larger, more temporally detailed data in the future and its effectiveness in governance, monitoring, and fairness should be tested in real-life scenarios.

LIST OF ABBREVIATIONS

AUTHORS’ CONTRIBUTIONS

Y.A. has contributed to study concept and design, data collection and data analysis. S.A. has contributed in writing the paper and results interpretation.

ETHICAL APPROVAL & INFORMED CONSENT

All procedures were carried out in accordance with institutional research ethics committee guidelines and Declaration of Helsinki. Informed consent was obtained from all participants. To ensure participant protection, all data were fully anonymized at the point of collection, and no personal or identifiable data was recorded.

AVAILABILITY OF DATA AND MATERIALS

The data will be made available on reasonable request by contacting the corresponding author [Y.A.]. Dataset: Kaggle – Credit Card Fraud Detection.

FUNDING

None.

CONFLICT OF INTEREST

The authors declare that there is no conflict of interest regarding the publication of this article.

ACKNOWLEDGEMENTS

The authors would also appreciate the help and support of the people who put together the open-source community and submitted the dataset: Credit Card Fraud Detection created in Kaggle, whose work made it clear and a replicated study possible. We also thank the individuals with whom we worked and reviewed who gave us technical knowledge at the modelling, design, test, and result verification stages. Google Colab with the libraries based on Python, such as Scikit-learn and SHAP, is used to conduct the computational experiments.

DECLARATION OF AI

During the preparation of this work the authors used ChatGPT for editing purposes. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.