Jialei Cao

Jialei Cao

Auto-generated profile, based on publicly available information

Jialei Cao

University of Pennsylvania, Philadelphia, PA, USA

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
8 July 2025 in IEEE Open Journal of the Computer Society
Verified
Journal Article
Open access
JCJialei Cao
Jialei Cao
WZWenxia Zheng
Wenxia Zheng
YGYao Ge
Yao Ge
JWJiyuan Wang
Jiyuan Wang
Financial fraud detection systems confront the persistent challenge of concept drift, where fraudulent patterns evolve continuously to evade detection mechanisms. Traditional rule-based methods and static machine learning models require frequent manual updates, failing to autonomously adapt to emerging fraud strategies. This paper presents DriftShield, a novel adaptive fraud detection framework that addresses these limitations through four key technical innovations: (1) the first application of Soft Actor-Critic (SAC) reinforcement learning with continuous action spaces to fraud detection, enabling simultaneous fine-grained optimization of detection thresholds and feature importance weights; (2) a dynamic feature reweighting mechanism that automatically adapts to evolving fraud patterns while providing interpretable insights into changing fraud strategies; (3) an adaptive experience replay buffer combining sliding windows with prioritized sampling to balance catastrophic forgetting prevention with rapid concept drift adaptation; and (4) an entropy-driven exploration framework with automatic temperature tuning that intelligently balances exploitation of known fraud patterns with discovery of emerging threats. Experimental evaluation demonstrates that DriftShield achieves 18% higher fraud detection rates while maintaining lower false positive rates compared to static models. The system demonstrates 57% faster adaptation times, recovering optimal performance within 280 transactions after significant concept drift compared to 650 transactions for the next-best reinforcement learning approach. DriftShield attains a cumulative detection rate of 0.849, representing a 7.7% improvement over existing methods and establishing the efficacy of continuous-action reinforcement learning for autonomous adaptation in dynamic adversarial environments.

19 May 2025 in IEEE Access
Verified
Journal Article
Open access
YTYutong Tan
Yutong Tan
BWBi Wu
Bi Wu
JCJialei Cao
Jialei Cao
BJBingying Jiang
Bingying Jiang
This paper introduces LLaMA-UTP, a novel framework for classifying and interpreting uncertain tax positions (UTPs) in technology companies. Tax professionals face significant challenges when analyzing UTPs: navigating vast bodies of intricate legal text, interpreting highly technical domain-specific language, maintaining accurate citations to authoritative sources, and producing justifications that comply with legal standards. We address these challenges with a knowledge-guided expert mixture system built upon a domain-adapted LLaMA architecture incorporating retrieval-augmented generation to ground analysis in relevant tax authorities, specialized expert subnetworks for different tax knowledge domains, and a tax citation knowledge graph that enhances interpretability through structured reasoning paths. Our comprehensive evaluation demonstrates that LLaMA-UTP achieves 90.5% accuracy in classifying UTP risk levels, significantly outperforming baseline models including LegalBERT (82.0%) and GPT-3.5 (78.4%). The system generates explanations with 93% citation accuracy and zero hallucinated references, as validated by tax professionals who rated its outputs as highly aligned with expert reasoning (4.2/5). Through ablation studies, we demonstrate that each architectural component contributes measurably to both accuracy and interpretability. This work makes significant contributions to both AI/ML methodology and tax law analysis by showcasing howknowledge-guided expert mixtures can enhance model performance in highly specialized professional domains while maintaining the explainability necessary for high-stakes legal applications.

19 November 2025 in IEEE Access
Verified
Journal Article
Open access
BJBingying Jiang
Bingying Jiang
BWBi Wu
Bi Wu
JCJialei Cao
Jialei Cao
YTYutong Tan
Yutong Tan
Fair value hierarchy classification under IFRS 13 presents a significant challenge in financial reporting, requiring accurate categorization of financial instruments into Level 1 (directly observable prices), Level 2 (indirectly observable inputs), or Level 3 (unobservable inputs). This classification critically impacts perceived valuation reliability, yet remains inconsistent when performed manually. We propose a novel hybrid architecture combining transformer-based attention with graph neural networks (GNNs) that jointly exploits temporal market patterns and inter-instrument relational dependencies. Our approach leverages transformer encoders to capture time-series dynamics in market data while utilizing graph neural networks to model structural relationships between instruments. To address limited labeled data availability, we implement self-supervised pretraining on unlabeled market data, enhancing model efficiency. The architecture’s attention mechanisms provide interpretable explanations crucial for audit and regulatory compliance. Experiments on open-source datasets of derivatives and structured products demonstrate that our model achieves 88% classification accuracy, significantly outperforming traditional methods including rule-based classifiers (73%) and XGBoost (80%). Notably, our model excels at identifying Level 3 instruments (F1-score 0.90) while maintaining interpretability through attention-based explanations that highlight key temporal periods and influential instrument relationships. Ablation studies confirm the complementary contributions of both transformer and GNN components. This work bridges advanced deep learning with financial compliance requirements, facilitating more consistent, accurate, and transparent valuation procedures that align with regulatory standards.