Selected work

Representative Papers

SHIELD: Multi-task Multi-distribution Vehicle Routing Solver with Sparsity and Hierarchy

Jun 10, 2025

To address the weak generalization capability of base models in multi-task, multi-distribution vehicle routing problems (MTMDVRP) and their difficulty adapting to heterogeneous customer distributions in real-world scenarios, this paper proposes the first unified solution framework. Methodologically, it innovatively integrates sparse computation (Mixture-of-Depths) with context-aware hierarchical clustering to establish a dual inductive bias mechanism, enabling adaptive representation learning across tasks and distributions; it further employs a deep decoder architecture to dynamically allocate computational resources and model spatial hierarchical structures. Extensive experiments across nine real-world maps and 144 VRP variants demonstrate that our approach significantly outperforms existing state-of-the-art methods, achieving substantial gains in generalization to unseen tasks and unknown distributions. This work establishes a scalable, robust, and general-purpose paradigm for complex real-world routing optimization.

1 citationsRead paper

NodeImport: Imbalanced Node Classification with Node Importance Assessment

Jul 15, 2026

This work addresses the challenge of class imbalance in graph node classification, where models tend to favor majority classes. To mitigate this bias, the authors propose a general-purpose framework that decouples node generation from selection. The method constructs a high-quality meta-set that approximates the global feature distribution, enabling efficient estimation of node importance. It dynamically selects labeled, unlabeled, and synthetically generated nodes whose inclusion yields significant gains in unbiased performance. Theoretical analysis supports the efficiency of the importance computation, and by integrating graph neural networks with meta-learning principles, the framework achieves fine-grained node selection. Extensive experiments demonstrate that the proposed approach consistently outperforms state-of-the-art methods across multiple datasets, effectively alleviating class imbalance and substantially improving overall classification performance.

0 citationsRead paper

Multi-Label Node Classification with Label Influence Propagation

Jul 01, 2026

This work addresses the challenge of effectively modeling complex inter-label dependencies in multi-label node classification on graph-structured data. The authors propose a novel approach that decouples the message-passing mechanism of graph neural networks into distinct propagation and transformation operations, enabling explicit analysis and quantification of positive and negative interactions among labels. Building upon this decomposition, they construct a label influence graph to propagate higher-order label effects and introduce a dynamic adjustment mechanism to optimize the learning process. To the best of our knowledge, this is the first method to systematically model inter-label influence relations in non-Euclidean graph data. Extensive experiments demonstrate that the proposed framework significantly outperforms state-of-the-art baselines across multiple benchmark datasets, achieving substantial improvements in multi-label node classification performance.

0 citationsRead paper

Combating Organized Platform Abuse: Amplifying Weak Risk Signals with Structural Information

May 08, 2026

This study addresses the inefficiency in detecting organized abuse—such as credit card fraud and promotional misuse—on large-scale online platforms. Grounded in the fraudster’s trilemma theory, the work reveals the structural invariance of “centralized cash-out” behavior and proposes a detection framework that is label-free, nearly parameter-free, white-box interpretable, and evasion-resistant. Relying solely on graph topology, the method aggregates low-precision individual weak signals into high-accuracy collective decisions through a linear-complexity (O(|E|)) statistical mechanism, without requiring seed users or domain-specific business logic. In real-world deployment, weak signals with only 16% precision are amplified to achieve over 91% precision and 99% recall, and the approach successfully detects cross-tier credit card fraud, demonstrating its robust generalizability across diverse attack patterns.

0 citationsRead paper

Efficient Densest Flow Queries in Transaction Flow Networks (Complete Version)

Feb 17, 2026

This work proposes the (S,T)-densest monetary flow (SDMF) query problem to effectively detect illicit activities such as money laundering and credit card fraud in transaction flow networks. Formally defined for the first time, the SDMF query seeks a subgraph that maximizes flow density while satisfying cardinality constraints over a given set of source nodes S and sink nodes T. Addressing this NP-hard problem, the authors design CONAN, a divide-and-conquer algorithm enhanced with approximate flow peeling and graph density optimization techniques to improve efficiency. Extensive experiments on large-scale real-world transaction networks—including Grab and NFT datasets—demonstrate that CONAN achieves up to three orders of magnitude speedup over baseline methods, substantially enhancing both query practicality and fraud detection capability.

0 citationsRead paper

Recent publications

Latest Papers

NodeImport: Imbalanced Node Classification with Node Importance Assessment

Jul 15, 2026

This work addresses the challenge of class imbalance in graph node classification, where models tend to favor majority classes. To mitigate this bias, the authors propose a general-purpose framework that decouples node generation from selection. The method constructs a high-quality meta-set that approximates the global feature distribution, enabling efficient estimation of node importance. It dynamically selects labeled, unlabeled, and synthetically generated nodes whose inclusion yields significant gains in unbiased performance. Theoretical analysis supports the efficiency of the importance computation, and by integrating graph neural networks with meta-learning principles, the framework achieves fine-grained node selection. Extensive experiments demonstrate that the proposed approach consistently outperforms state-of-the-art methods across multiple datasets, effectively alleviating class imbalance and substantially improving overall classification performance.

0 citationsRead paper

Multi-Label Node Classification with Label Influence Propagation

Jul 01, 2026

This work addresses the challenge of effectively modeling complex inter-label dependencies in multi-label node classification on graph-structured data. The authors propose a novel approach that decouples the message-passing mechanism of graph neural networks into distinct propagation and transformation operations, enabling explicit analysis and quantification of positive and negative interactions among labels. Building upon this decomposition, they construct a label influence graph to propagate higher-order label effects and introduce a dynamic adjustment mechanism to optimize the learning process. To the best of our knowledge, this is the first method to systematically model inter-label influence relations in non-Euclidean graph data. Extensive experiments demonstrate that the proposed framework significantly outperforms state-of-the-art baselines across multiple benchmark datasets, achieving substantial improvements in multi-label node classification performance.

0 citationsRead paper

Combating Organized Platform Abuse: Amplifying Weak Risk Signals with Structural Information

May 08, 2026

This study addresses the inefficiency in detecting organized abuse—such as credit card fraud and promotional misuse—on large-scale online platforms. Grounded in the fraudster’s trilemma theory, the work reveals the structural invariance of “centralized cash-out” behavior and proposes a detection framework that is label-free, nearly parameter-free, white-box interpretable, and evasion-resistant. Relying solely on graph topology, the method aggregates low-precision individual weak signals into high-accuracy collective decisions through a linear-complexity (O(|E|)) statistical mechanism, without requiring seed users or domain-specific business logic. In real-world deployment, weak signals with only 16% precision are amplified to achieve over 91% precision and 99% recall, and the approach successfully detects cross-tier credit card fraud, demonstrating its robust generalizability across diverse attack patterns.

0 citationsRead paper

Efficient Densest Flow Queries in Transaction Flow Networks (Complete Version)

Feb 17, 2026

This work proposes the (S,T)-densest monetary flow (SDMF) query problem to effectively detect illicit activities such as money laundering and credit card fraud in transaction flow networks. Formally defined for the first time, the SDMF query seeks a subgraph that maximizes flow density while satisfying cardinality constraints over a given set of source nodes S and sink nodes T. Addressing this NP-hard problem, the authors design CONAN, a divide-and-conquer algorithm enhanced with approximate flow peeling and graph density optimization techniques to improve efficiency. Extensive experiments on large-scale real-world transaction networks—including Grab and NFT datasets—demonstrate that CONAN achieves up to three orders of magnitude speedup over baseline methods, substantially enhancing both query practicality and fraud detection capability.

0 citationsRead paper

Deep Learning Option Pricing with Market Implied Volatility Surfaces

Sep 07, 2025

This paper addresses the efficient and accurate pricing of American put and arithmetic Asian options by proposing an end-to-end deep learning framework grounded in the market-implied volatility surface. Methodologically, a variational autoencoder (VAE) compresses the high-dimensional volatility surface into a low-dimensional latent representation, which is then mapped directly to option prices via a multilayer perceptron (MLP), jointly optimizing surface representation and pricing. Training data are generated using QuantLib, with a staged training and full fine-tuning strategy employed. Compared to conventional numerical methods, the model achieves order-of-magnitude speedup via single forward-pass inference while maintaining high accuracy across diverse contracts—only exhibiting minor errors for long-maturity at-the-money options. The key contribution lies in the first unified, scalable, and data-adaptive end-to-end neural architecture that jointly learns volatility surface dimensionality reduction and option pricing, enabling rapid valuation of arbitrary path-dependent derivatives.

0 citationsRead paper