Selected work

Representative Papers

Causal Label Recovery in Payment Networks

May 27, 2026

This study addresses systematic biases in fraud detection models within payment networks arising from chargeback labels, including authorization interception, issuer underreporting, delayed settlement, and label contamination. The authors formulate label generation as a sequential missing-data problem with a three-stage propensity scoring mechanism and an explicit contamination layer. They propose the Sequential Triple-Robust (STR) estimator—the first method capable of simultaneously correcting all four bias sources. By decoupling model training from the chargeback maturation cycle, STR enables effective use of data just days old. The estimator achieves strictly lower mean squared error than naive approaches at any sample size, and provides theoretically grounded guidance for optimal training window selection, finite-sample confidence intervals, and formal statistical guarantees.

0 citationsRead paper

The Fundamental Limits of Fraud Detection in Card Payment Networks

May 26, 2026

This study reveals that the long-standing performance limitations in bank card payment fraud detection stem fundamentally from inherent information deficiencies within the payment ecosystem, rather than insufficient model capacity. By formulating card authorization as a sequential decision-making problem characterized by delayed, censored, corrupted, and counterfactually missing feedback, the work establishes the first learnability lower bound for this setting. Through minimax regret analysis grounded in online learning and counterfactual reasoning—without requiring access to real transaction data—it demonstrates that information quality, particularly in issuer reporting and dispute resolution, constitutes the primary bottleneck, with heterogeneity further exacerbating learning difficulty. The theoretical analysis shows that the regret lower bound scales multiplicatively with the rate of information deficiency, indicating that improving reporting fidelity offers a more effective path to breaking through performance ceilings than increasing model complexity.

0 citationsRead paper

Retrieval Augmented Generation (RAG) for Fintech: Agentic Design and Evaluation

Oct 29, 2025

In financial technology domains, dense terminology and pervasive acronyms degrade retrieval and generation performance in Retrieval-Augmented Generation (RAG) systems. To address this, we propose a modular multi-agent collaborative RAG framework that integrates context-aware acronym resolution, keyword-guided iterative subquery decomposition, intelligent query rewriting, and cross-encoder re-ranking—enabling end-to-end precise retrieval and generation of domain-specific knowledge. Evaluated on an enterprise-scale financial knowledge base across 85 expert-curated question-answer pairs, our framework significantly outperforms conventional RAG baselines: retrieval accuracy improves by 23.6%, and relevance scores increase by 19.4%. These results validate the efficacy of the multi-agent collaboration paradigm in highly specialized domains. Although introducing moderate latency, the substantial gains in accuracy demonstrably outweigh the computational overhead.

0 citationsRead paper

ScaleCall - Agentic Tool Calling at Scale for Fintech: Challenges, Methods, and Deployment Insights

Oct 29, 2025

In highly regulated domains such as fintech, LLM-based tool invocation faces challenges including on-premises deployment requirements, stringent compliance constraints, and ambiguity arising from functional overlaps among APIs. To address these, this paper proposes ScaleCall—a production-ready framework for enterprise-grade API orchestration and automated data engineering. Its core innovation is a flexible hybrid tool retrieval architecture integrating embedding-based retrieval, prompt-driven ranking, and collaborative re-ranking. Empirical analysis demonstrates that domain-specific characteristics exert a stronger influence on retrieval performance than algorithmic choices alone. Evaluated in Mastercard’s real-world, regulated environment, embedding-based retrieval achieves low latency and superior ambiguity resolution in large-scale tool repositories, while the hybrid strategy shows promise in high-precision scenarios. This work establishes a deployable paradigm and a systematic evaluation benchmark for LLM tool invocation in compliance-sensitive settings.

0 citationsRead paper

Flow Matters: Directional and Expressive GNNs for Heterophilic Graphs

Aug 31, 2025

Traditional GNNs suffer from performance degradation on heterophilic graphs due to their reliance on the local homophily assumption. To address this, this paper proposes a novel message-passing paradigm that jointly incorporates edge directionality and high-order expressivity. We introduce two models: Poly, which employs learnable permutation-equivariant higher-order polynomial aggregation, and Dir-Poly, a direction-aware variant that explicitly distinguishes incoming and outgoing edges. This work is the first to synergistically model directional inductive bias and polynomial expressivity, revealing their complementary roles in feature learning for heterophilic graphs. Both models retain linear time complexity and exhibit strong scalability. Extensive experiments on five standard heterophilic graph benchmarks demonstrate that Poly significantly outperforms existing methods; Dir-Poly further improves performance on naturally directed graphs—validating the effectiveness and generalizability of our joint modeling approach.

0 citationsRead paper

Recent publications

Latest Papers

Causal Label Recovery in Payment Networks

May 27, 2026

This study addresses systematic biases in fraud detection models within payment networks arising from chargeback labels, including authorization interception, issuer underreporting, delayed settlement, and label contamination. The authors formulate label generation as a sequential missing-data problem with a three-stage propensity scoring mechanism and an explicit contamination layer. They propose the Sequential Triple-Robust (STR) estimator—the first method capable of simultaneously correcting all four bias sources. By decoupling model training from the chargeback maturation cycle, STR enables effective use of data just days old. The estimator achieves strictly lower mean squared error than naive approaches at any sample size, and provides theoretically grounded guidance for optimal training window selection, finite-sample confidence intervals, and formal statistical guarantees.

0 citationsRead paper

The Fundamental Limits of Fraud Detection in Card Payment Networks

May 26, 2026

This study reveals that the long-standing performance limitations in bank card payment fraud detection stem fundamentally from inherent information deficiencies within the payment ecosystem, rather than insufficient model capacity. By formulating card authorization as a sequential decision-making problem characterized by delayed, censored, corrupted, and counterfactually missing feedback, the work establishes the first learnability lower bound for this setting. Through minimax regret analysis grounded in online learning and counterfactual reasoning—without requiring access to real transaction data—it demonstrates that information quality, particularly in issuer reporting and dispute resolution, constitutes the primary bottleneck, with heterogeneity further exacerbating learning difficulty. The theoretical analysis shows that the regret lower bound scales multiplicatively with the rate of information deficiency, indicating that improving reporting fidelity offers a more effective path to breaking through performance ceilings than increasing model complexity.

0 citationsRead paper

Retrieval Augmented Generation (RAG) for Fintech: Agentic Design and Evaluation

Oct 29, 2025

In financial technology domains, dense terminology and pervasive acronyms degrade retrieval and generation performance in Retrieval-Augmented Generation (RAG) systems. To address this, we propose a modular multi-agent collaborative RAG framework that integrates context-aware acronym resolution, keyword-guided iterative subquery decomposition, intelligent query rewriting, and cross-encoder re-ranking—enabling end-to-end precise retrieval and generation of domain-specific knowledge. Evaluated on an enterprise-scale financial knowledge base across 85 expert-curated question-answer pairs, our framework significantly outperforms conventional RAG baselines: retrieval accuracy improves by 23.6%, and relevance scores increase by 19.4%. These results validate the efficacy of the multi-agent collaboration paradigm in highly specialized domains. Although introducing moderate latency, the substantial gains in accuracy demonstrably outweigh the computational overhead.

0 citationsRead paper

ScaleCall - Agentic Tool Calling at Scale for Fintech: Challenges, Methods, and Deployment Insights

Oct 29, 2025

In highly regulated domains such as fintech, LLM-based tool invocation faces challenges including on-premises deployment requirements, stringent compliance constraints, and ambiguity arising from functional overlaps among APIs. To address these, this paper proposes ScaleCall—a production-ready framework for enterprise-grade API orchestration and automated data engineering. Its core innovation is a flexible hybrid tool retrieval architecture integrating embedding-based retrieval, prompt-driven ranking, and collaborative re-ranking. Empirical analysis demonstrates that domain-specific characteristics exert a stronger influence on retrieval performance than algorithmic choices alone. Evaluated in Mastercard’s real-world, regulated environment, embedding-based retrieval achieves low latency and superior ambiguity resolution in large-scale tool repositories, while the hybrid strategy shows promise in high-precision scenarios. This work establishes a deployable paradigm and a systematic evaluation benchmark for LLM tool invocation in compliance-sensitive settings.

0 citationsRead paper

Flow Matters: Directional and Expressive GNNs for Heterophilic Graphs

Aug 31, 2025

Traditional GNNs suffer from performance degradation on heterophilic graphs due to their reliance on the local homophily assumption. To address this, this paper proposes a novel message-passing paradigm that jointly incorporates edge directionality and high-order expressivity. We introduce two models: Poly, which employs learnable permutation-equivariant higher-order polynomial aggregation, and Dir-Poly, a direction-aware variant that explicitly distinguishes incoming and outgoing edges. This work is the first to synergistically model directional inductive bias and polynomial expressivity, revealing their complementary roles in feature learning for heterophilic graphs. Both models retain linear time complexity and exhibit strong scalability. Extensive experiments on five standard heterophilic graph benchmarks demonstrate that Poly significantly outperforms existing methods; Dir-Poly further improves performance on naturally directed graphs—validating the effectiveness and generalizability of our joint modeling approach.

0 citationsRead paper