Xu Han
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xu Han has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
Jingyun Yang
2 shared publications- Xuguang Zhang
University of Gloucestershire, Cheltenham, U.K.
2 shared publications Huijie Fan
Hunan Agricultural University, Changsha, China
1 shared publication- Jiaying Chen
SC Johnson Graduate School of Management, Cornell University, Ithaca, NY 14853, USA
1 shared publication Pan Li
Business School, University of Hull, Hull, U.K.
1 shared publication- Wenjie Ping
Georgetown University, Washington, DC, USA
1 shared publication
Latest Publications
- Open Access
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Change-Point-Aware Lifecycle Validation of Credit Risk Models Under Economic Regime Shifts
Credit risk models deployed across economic regime shifts face two coupled problems: their probability-of-default (PD) estimates drift, and the outcomes needed to verify them arrive only after the performance window has closed. This paper proposes CP-LV, a change-point-aware lifecycle validation framework that decides each month whether a PD model should be kept, recalibrated or revised. CP-LV fuses three evidence channels, namely a re-centring CUSUM on a FRED-MD macro-stress index, a population-stability channel and a maturity-adjusted riskadjusted CUSUM (MA-RA-CUSUM) that extracts calibration evidence from partially matured loans, and resolves alarms with a two-stage closed test that learns structural revisions only from fully matured cohorts while realigning the PD level with fresh, maturity-adjusted evidence. We prove that the MA-RA-CUSUM is an exact sequential decomposition of the matured-label likelihood ratio and make it robust to shifts in early-payment defaults through left truncation and risk-banded emergence curves. On 68,927 Prosper loans originated between 2006 and 2013, a period spanning the 2008 crisis, a platform suspension and a relaunch, CP-LV achieved the highest AUROC (0.697) and the lowest Brier score (0.0687), expected calibration error (0.0082, 45% below a static model) and misclassification cost among eight lifecycle policies, with 70% fewer full refits than quarterly retraining. In a Monte-Carlo benchmark with known change points, the outcome channel detected shifts 1-4 months earlier than matured-label monitoring, and CP-LV attained the lowest post-change calibration error in four of five shift scenarios.
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessChange-Point-Aware Lifecycle Validation of Credit Risk Models Under Economic Regime Shifts
Credit risk models deployed across economic regime shifts face two coupled problems: their probability-of-default (PD) estimates drift, and the outcomes needed to verify them arrive only after the performance window has closed. This paper proposes CP-LV, a change-point-aware lifecycle validation framework that decides each month whether a PD model should be kept, recalibrated or revised. CP-LV fuses three evidence channels, namely a re-centring CUSUM on a FRED-MD macro-stress index, a population-stability channel and a maturity-adjusted riskadjusted CUSUM (MA-RA-CUSUM) that extracts calibration evidence from partially matured loans, and resolves alarms with a two-stage closed test that learns structural revisions only from fully matured cohorts while realigning the PD level with fresh, maturity-adjusted evidence. We prove that the MA-RA-CUSUM is an exact sequential decomposition of the matured-label likelihood ratio and make it robust to shifts in early-payment defaults through left truncation and risk-banded emergence curves. On 68,927 Prosper loans originated between 2006 and 2013, a period spanning the 2008 crisis, a platform suspension and a relaunch, CP-LV achieved the highest AUROC (0.697) and the lowest Brier score (0.0687), expected calibration error (0.0082, 45% below a static model) and misclassification cost among eight lifecycle policies, with 70% fewer full refits than quarterly retraining. In a Monte-Carlo benchmark with known change points, the outcome channel detected shifts 1-4 months earlier than matured-label monitoring, and CP-LV attained the lowest post-change calibration error in four of five shift scenarios.
- Open Access
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Drift-Resilient Explainable Graph Learning for Financial Fraud Detection: Joint Evaluation of Risk Calibration and Explanation Stability
Graph neural networks detect organized payment fraud by exploiting shared cards, devices and e-mail accounts, yet deployed detectors face heterogeneous drift that erodes not only accuracy but also the calibration of risk scores and the stability of explanations. This paper presents a drift-resilient explainable graph-learning framework that evaluates and maintains these three properties jointly. A GraphSAGE detector is equipped with exact gradient×input attributions computed by forward-mode differentiation, which separate a transaction's own evidence from evidence propagated through the graph. A label-free monitor standardizes five statistics, namely feature, score, signed prior, decision-focused attribution and structural-reliance drift, against a reference null, and a Diagnose-Then-Act (DTA) policy maps them to the cheapest sufficient action: no action, label-free EM recalibration, delayed-label Platt refresh, or change-point-aware retraining. On FinDrift-G, an IEEE-CIS-style transaction-graph benchmark with ground-truth drift types and fraud mechanisms, each drift type leaves a distinct fingerprint: benign covariate drift drives the feature-PSI z-score to 1073 while PR-AUC stays at 0.909, whereas account-takeover drift reduces the recall of a static GraphSAGE from 0.89 to 0.13. Under harmful drift, DTA reaches a PR-AUC of 0.843 against 0.845 for retraining at every step while using 2.0 instead of 22.3 retraining events per stream, attains the lowest expected calibration error (0.77%), and it retrains only 1.0 times per stream when drift is absent or benign. Discrimination, calibration and explanation stability therefore need to be monitored together, and diagnosing the drift type, rather than merely detecting drift, determines the right maintenance action.
- 13 November 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDrift-Resilient Explainable Graph Learning for Financial Fraud Detection: Joint Evaluation of Risk Calibration and Explanation Stability
Graph neural networks detect organized payment fraud by exploiting shared cards, devices and e-mail accounts, yet deployed detectors face heterogeneous drift that erodes not only accuracy but also the calibration of risk scores and the stability of explanations. This paper presents a drift-resilient explainable graph-learning framework that evaluates and maintains these three properties jointly. A GraphSAGE detector is equipped with exact gradient×input attributions computed by forward-mode differentiation, which separate a transaction's own evidence from evidence propagated through the graph. A label-free monitor standardizes five statistics, namely feature, score, signed prior, decision-focused attribution and structural-reliance drift, against a reference null, and a Diagnose-Then-Act (DTA) policy maps them to the cheapest sufficient action: no action, label-free EM recalibration, delayed-label Platt refresh, or change-point-aware retraining. On FinDrift-G, an IEEE-CIS-style transaction-graph benchmark with ground-truth drift types and fraud mechanisms, each drift type leaves a distinct fingerprint: benign covariate drift drives the feature-PSI z-score to 1073 while PR-AUC stays at 0.909, whereas account-takeover drift reduces the recall of a static GraphSAGE from 0.89 to 0.13. Under harmful drift, DTA reaches a PR-AUC of 0.843 against 0.845 for retraining at every step while using 2.0 instead of 22.3 retraining events per stream, attains the lowest expected calibration error (0.77%), and it retrains only 1.0 times per stream when drift is absent or benign. Discrimination, calibration and explanation stability therefore need to be monitored together, and diagnosing the drift type, rather than merely detecting drift, determines the right maintenance action.
- Open Access
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Explainable Artificial Intelligence for Dynamic Credit Risk Assessment: X-DyAR, an Additive Temporal Residual Fusion Framework for Financial Institutions
Credit risk models used at loan origination are typically built from applicant-level attributes observed at a single decision point, while economic conditions evolve and can alter the background probability of default. Adding external time series to a strong tabular model is technically non-trivial in regulated settings: naive fusion can introduce temporal leakage, opaque crosssource interactions can make explanations difficult to audit, and an external data outage can destabilize production predictions. This paper proposes X-DyAR (Explainable Dynamic Additive Residual Fusion), a concrete hybrid deep-learning method that augments a CatBoost applicationtime backbone with a gated recurrent unit (GRU) that learns a month-level residual log-odds adjustment from lagged macroeconomic trajectories. The architecture is additive by construction: the final logit equals the static CatBoost logit plus one dynamic source contribution. TreeSHAP decomposes the static component into applicant features, while the dynamic residual is an exact source-level contribution; their sum reconstructs the prediction to floating-point precision. A source-availability gate and source-level temporal dropout provide an exact fallback to the static model when the external source is entirely unavailable. We evaluate the method on a real Lending Club granting dataset with 1,347,681 loans and a common leakage-controlled cohort of 1,335,335 loans, fused with monthly U.S. unemployment, policy-rate, inflation, 10-year Treasury-yield, and yield-slope signals. Training uses loans through 2015, validation uses 2016, and all 225,277 loans from 2017-2018 form the out-of-time test. The best six-month X-DyAR model attains AUROC 0.68329, AUPRC 0.35520, Brier 0.15586, log loss 0.48209, and 15-bin ECE 0.01109, versus 0.68194, 0.35298, 0.15790, 0.48825, and 0.04109 for the strong static CatBoost baseline. The AUROC gain is deliberately reported as small rather than overstated; a paired bootstrap using 100 resamples of 25,000 test loans yields an empirical AUROC-difference interval of [0.00053, 0.00225]. Ablations, 3/6/12-month sensitivity, source-missingness stress tests, year-wise analysis, FICO-stratified analysis, and exact explanation-completeness checks show that the main benefit is a transparent, robust dynamic calibration layer rather than a dramatic within-year ranking improvement. The results support additive temporal residual fusion as a practical design pattern when financial institutions need dynamic probability adjustment without surrendering auditability or safe fallback behavior.
- 19 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessExplainable Artificial Intelligence for Dynamic Credit Risk Assessment: X-DyAR, an Additive Temporal Residual Fusion Framework for Financial Institutions
Credit risk models used at loan origination are typically built from applicant-level attributes observed at a single decision point, while economic conditions evolve and can alter the background probability of default. Adding external time series to a strong tabular model is technically non-trivial in regulated settings: naive fusion can introduce temporal leakage, opaque crosssource interactions can make explanations difficult to audit, and an external data outage can destabilize production predictions. This paper proposes X-DyAR (Explainable Dynamic Additive Residual Fusion), a concrete hybrid deep-learning method that augments a CatBoost applicationtime backbone with a gated recurrent unit (GRU) that learns a month-level residual log-odds adjustment from lagged macroeconomic trajectories. The architecture is additive by construction: the final logit equals the static CatBoost logit plus one dynamic source contribution. TreeSHAP decomposes the static component into applicant features, while the dynamic residual is an exact source-level contribution; their sum reconstructs the prediction to floating-point precision. A source-availability gate and source-level temporal dropout provide an exact fallback to the static model when the external source is entirely unavailable. We evaluate the method on a real Lending Club granting dataset with 1,347,681 loans and a common leakage-controlled cohort of 1,335,335 loans, fused with monthly U.S. unemployment, policy-rate, inflation, 10-year Treasury-yield, and yield-slope signals. Training uses loans through 2015, validation uses 2016, and all 225,277 loans from 2017-2018 form the out-of-time test. The best six-month X-DyAR model attains AUROC 0.68329, AUPRC 0.35520, Brier 0.15586, log loss 0.48209, and 15-bin ECE 0.01109, versus 0.68194, 0.35298, 0.15790, 0.48825, and 0.04109 for the strong static CatBoost baseline. The AUROC gain is deliberately reported as small rather than overstated; a paired bootstrap using 100 resamples of 25,000 test loans yields an empirical AUROC-difference interval of [0.00053, 0.00225]. Ablations, 3/6/12-month sensitivity, source-missingness stress tests, year-wise analysis, FICO-stratified analysis, and exact explanation-completeness checks show that the main benefit is a transparent, robust dynamic calibration layer rather than a dramatic within-year ranking improvement. The results support additive temporal residual fusion as a practical design pattern when financial institutions need dynamic probability adjustment without surrendering auditability or safe fallback behavior.
- Top Cited PapersOpen Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 1
Adaptive Event-Driven Labeling: A Neuro-Symbolic Multiagent Framework for Causal Inference in Non-Stationary Time Series
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 1
Top Cited PapersOpen AccessAdaptive Event-Driven Labeling: A Neuro-Symbolic Multiagent Framework for Causal Inference in Non-Stationary Time Series
- Top Cited PapersOpen Access
- 18 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 6, 426-437
FraudGNN-RL: A Graph Neural Network With Reinforcement Learning for Adaptive Financial Fraud Detection
As financial systems become increasingly complex and interconnected, traditional fraud detection methods struggle to keep pace with sophisticated fraudulent activities. This paper introduces FraudGNN-RL, an innovative framework that combines Graph Neural Networks (GNNs) with Reinforcement Learning (RL) for adaptive and context-aware financial fraud detection. Our approach models financial transactions as a dynamic graph, where entities (e.g., users, merchants) are nodes and transactions form edges. We propose a novel GNN architecture, Temporal-Spatial-Semantic Graph Convolution (TSSGC), which simultaneously captures temporal patterns, spatial relationships, and semantic information in transaction data. The RL component, implemented as a Deep Q-Network (DQN), dynamically adjusts the fraud detection threshold and feature importance, allowing the model to adapt to evolving fraud patterns and minimize detection costs. We further introduce a Federated Learning scheme to enable collaborative model training across multiple financial institutions while preserving data privacy. Extensive experiments on a large-scale, real-world financial dataset demonstrate that FraudGNN-RL outperforms state-of-the-art baselines, achieving a 97.3% F1-score and reducing false positives by 31% compared to the best-performing baseline. Our framework also shows remarkable resilience to concept drift and adversarial attacks, maintaining high performance over extended periods. These results suggest that FraudGNN-RL offers a robust, adaptive, and privacy-preserving solution for financial fraud detection in the era of Big Data and interconnected financial ecosystems.
- 18 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of the Computer Society
- Vol. 6, 426-437
Top Cited PapersOpen AccessFraudGNN-RL: A Graph Neural Network With Reinforcement Learning for Adaptive Financial Fraud Detection
As financial systems become increasingly complex and interconnected, traditional fraud detection methods struggle to keep pace with sophisticated fraudulent activities. This paper introduces FraudGNN-RL, an innovative framework that combines Graph Neural Networks (GNNs) with Reinforcement Learning (RL) for adaptive and context-aware financial fraud detection. Our approach models financial transactions as a dynamic graph, where entities (e.g., users, merchants) are nodes and transactions form edges. We propose a novel GNN architecture, Temporal-Spatial-Semantic Graph Convolution (TSSGC), which simultaneously captures temporal patterns, spatial relationships, and semantic information in transaction data. The RL component, implemented as a Deep Q-Network (DQN), dynamically adjusts the fraud detection threshold and feature importance, allowing the model to adapt to evolving fraud patterns and minimize detection costs. We further introduce a Federated Learning scheme to enable collaborative model training across multiple financial institutions while preserving data privacy. Extensive experiments on a large-scale, real-world financial dataset demonstrate that FraudGNN-RL outperforms state-of-the-art baselines, achieving a 97.3% F1-score and reducing false positives by 31% compared to the best-performing baseline. Our framework also shows remarkable resilience to concept drift and adversarial attacks, maintaining high performance over extended periods. These results suggest that FraudGNN-RL offers a robust, adaptive, and privacy-preserving solution for financial fraud detection in the era of Big Data and interconnected financial ecosystems.
- Open Access
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 190079-190091
Enhancing Time Series Product Demand Forecasting With Hybrid Attention-Based Deep Learning Models
Time series forecasting plays a crucial role in various industries, particularly in predicting product demand for effective supply chain management. This paper presents a novel approach to time series forecasting by leveraging advanced deep learning techniques, specifically focusing on hybrid models that combine attention mechanisms with traditional recurrent neural networks. Our proposed method, the Hybrid Attention-based Long Short-Term Memory (HA-LSTM) network, integrates multi-head self-attention modules with LSTM layers to capture both long-term dependencies and local temporal patterns in time series data. We evaluate our model on a large-scale retail dataset from the "Predict Future Sales" competition, demonstrating its effectiveness in handling complex, real-world time series with multiple seasonalities and trends. Experimental results show that the HA-LSTM outperforms state-of-the-art baselines, including ARIMA, Prophet, and vanilla LSTM models, achieving a 15% improvement in Mean Absolute Percentage Error (MAPE) and a 12% reduction in Root Mean Square Error (RMSE). Furthermore, we conduct an extensive ablation study to analyze the contribution of each component in our hybrid model. The proposed approach not only enhances forecasting accuracy but also provides interpretable attention weights, offering insights into the relative importance of different time steps in making predictions. This research contributes to the growing body of work on deep learning for time series analysis and offers practical implications for improving demand forecasting in retail and supply chain management.
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 190079-190091
Open AccessEnhancing Time Series Product Demand Forecasting With Hybrid Attention-Based Deep Learning Models
Time series forecasting plays a crucial role in various industries, particularly in predicting product demand for effective supply chain management. This paper presents a novel approach to time series forecasting by leveraging advanced deep learning techniques, specifically focusing on hybrid models that combine attention mechanisms with traditional recurrent neural networks. Our proposed method, the Hybrid Attention-based Long Short-Term Memory (HA-LSTM) network, integrates multi-head self-attention modules with LSTM layers to capture both long-term dependencies and local temporal patterns in time series data. We evaluate our model on a large-scale retail dataset from the "Predict Future Sales" competition, demonstrating its effectiveness in handling complex, real-world time series with multiple seasonalities and trends. Experimental results show that the HA-LSTM outperforms state-of-the-art baselines, including ARIMA, Prophet, and vanilla LSTM models, achieving a 15% improvement in Mean Absolute Percentage Error (MAPE) and a 12% reduction in Root Mean Square Error (RMSE). Furthermore, we conduct an extensive ablation study to analyze the contribution of each component in our hybrid model. The proposed approach not only enhances forecasting accuracy but also provides interpretable attention weights, offering insights into the relative importance of different time steps in making predictions. This research contributes to the growing body of work on deep learning for time series analysis and offers practical implications for improving demand forecasting in retail and supply chain management.