Yang Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yang Liu has a total of 49 co-authors. Here are the top co-authors by number of shared publications:
- Chuan Zhou
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China
3 shared publications - Xixun Lin
Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China
3 shared publications - Zhao Li
Zhejiang Lab
3 shared publications - Bin Chong
Peking University, China
2 shared publications - Hongyang Chen
Zhejiang Lab
2 shared publications - Lixin Zou
School of Cyber Science and Engineering, Wuhan University, Wuhan, China
2 shared publications
Latest Publications
- Open Access
- 21 October 2026
- Published by MDPI AG in Preprints.org
Data–Mechanism Fusion for Physical Intelligence: Principles, Models and Agents
Physical prediction, inference and control combine incomplete mechanistic models with limited observations. We argue that data–mechanism fusion should be designed around the target quantity or downstream decision. This Perspective distinguishes learning unresolved terms within mechanistic models, structuring learning with physical relations, and correcting predictions or updating estimates through observations. Constrained-estimation examples show that a misspecified constraint can reduce total estimation risk while increasing target-specific risk. Oscillator and force-limited lifting examples illustrate how excitation and calibration resolve decision-relevant uncertainty, and how a probe's information, cost and effect on subsequent motion determine its value. Selected studies in weather forecasting, atomistic modelling and robotics illustrate these roles in practice. We use Physics-Native Foundation Model (PNFM) as an operational term for a pretrained model that reuses learned components across a declared family of tasks or physical conditions, with specified physical structure entering its computation. We propose evaluating reuse through supported queries, task risk, adaptation and revalidation costs, and retained performance, with mechanism-based, data-driven and fused comparators. Physical structure, learned quantities and evidence acquisition should therefore be chosen together according to the task they support.
- 21 October 2026
- Published by MDPI AG in Preprints.org
Open AccessData–Mechanism Fusion for Physical Intelligence: Principles, Models and Agents
Physical prediction, inference and control combine incomplete mechanistic models with limited observations. We argue that data–mechanism fusion should be designed around the target quantity or downstream decision. This Perspective distinguishes learning unresolved terms within mechanistic models, structuring learning with physical relations, and correcting predictions or updating estimates through observations. Constrained-estimation examples show that a misspecified constraint can reduce total estimation risk while increasing target-specific risk. Oscillator and force-limited lifting examples illustrate how excitation and calibration resolve decision-relevant uncertainty, and how a probe's information, cost and effect on subsequent motion determine its value. Selected studies in weather forecasting, atomistic modelling and robotics illustrate these roles in practice. We use Physics-Native Foundation Model (PNFM) as an operational term for a pretrained model that reuses learned components across a declared family of tasks or physical conditions, with specified physical structure entering its computation. We propose evaluating reuse through supported queries, task risk, adaptation and revalidation costs, and retained performance, with mechanism-based, data-driven and fused comparators. Physical structure, learned quantities and evidence acquisition should therefore be chosen together according to the task they support.
- Open Access
- 14 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Knowledge Discovery From Data
Zero-Shot and Few-Shot Hydrogen Demand Forecasting with Time Series Foundation Models
Accurate hydrogen demand forecasting is essential for optimizing infrastructure investments in the emerging hydrogen economy, yet the scarcity of historical consumption data poses a fundamental challenge for conventional time series methods. We present a systematic evaluation of time series foundation models for hydrogen demand prediction, and propose DEFA (Data-Efficient Foundation model Adaptation), a two-stage framework designed to maximize few-shot adaptation performance under extreme data scarcity. DEFA first employs uncertainty-guided active sample selection, which leverages the foundation model's own zero-shot predictive variance to identify the most informative training sequences while maintaining diversity in embedding space. It then applies augmentation-enhanced consistency fine-tuning, combining temporal data augmentation with a consistency regularization loss that penalizes divergent predictions on semantically equivalent inputs, thereby mitigating overfitting when labeled data is scarce. We evaluate five pre-trained foundation models on two publicly available hydrogen datasets. Our experiments show that foundation models achieve strong zero-shot performance without domain-specific training, and that DEFA substantially improves few-shot adaptation efficiency, effectively halving the data requirement for a given accuracy target compared to standard random-sampling fine-tuning. Probabilistic foundation models further provide well-calibrated uncertainty estimates, enabling risk-aware infrastructure planning. To establish that these findings are not specific to the two hydrogen datasets, we further validate DEFA on six standard forecasting benchmarks (ETTh1, ETTh2, ETTm1, Electricity, Traffic, and M5 retail), where it delivers consistent 8.0–11.6% MAE reductions; we benchmark our selection rule against established active-learning strategies, extend the zero-shot comparison to recent foundation models, and provide a theoretical result linking the consistency objective to variance control in the few-shot regime. Our findings establish foundation models, enhanced by data-efficient adaptation, as a viable and practical solution for hydrogen demand forecasting in data-scarce environments.
- 14 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Knowledge Discovery From Data
Open AccessZero-Shot and Few-Shot Hydrogen Demand Forecasting with Time Series Foundation Models
Accurate hydrogen demand forecasting is essential for optimizing infrastructure investments in the emerging hydrogen economy, yet the scarcity of historical consumption data poses a fundamental challenge for conventional time series methods. We present a systematic evaluation of time series foundation models for hydrogen demand prediction, and propose DEFA (Data-Efficient Foundation model Adaptation), a two-stage framework designed to maximize few-shot adaptation performance under extreme data scarcity. DEFA first employs uncertainty-guided active sample selection, which leverages the foundation model's own zero-shot predictive variance to identify the most informative training sequences while maintaining diversity in embedding space. It then applies augmentation-enhanced consistency fine-tuning, combining temporal data augmentation with a consistency regularization loss that penalizes divergent predictions on semantically equivalent inputs, thereby mitigating overfitting when labeled data is scarce. We evaluate five pre-trained foundation models on two publicly available hydrogen datasets. Our experiments show that foundation models achieve strong zero-shot performance without domain-specific training, and that DEFA substantially improves few-shot adaptation efficiency, effectively halving the data requirement for a given accuracy target compared to standard random-sampling fine-tuning. Probabilistic foundation models further provide well-calibrated uncertainty estimates, enabling risk-aware infrastructure planning. To establish that these findings are not specific to the two hydrogen datasets, we further validate DEFA on six standard forecasting benchmarks (ETTh1, ETTh2, ETTm1, Electricity, Traffic, and M5 retail), where it delivers consistent 8.0–11.6% MAE reductions; we benchmark our selection rule against established active-learning strategies, extend the zero-shot comparison to recent foundation models, and provide a theoretical result linking the consistency objective to variance control in the few-shot regime. Our findings establish foundation models, enhanced by data-efficient adaptation, as a viable and practical solution for hydrogen demand forecasting in data-scarce environments.
- Open Access
- 10 October 2026
- Published by MDPI AG in Preprints.org
ScoAI: Toward OR-Native Decision Intelligence
Foundation models and AI agents are increasingly used to read operational requests, estimate uncertain quantities, and draft plans. In supply chains, these outputs feed decisions that must respect costs, stock and capacity limits, and business rules, and whose effects become visible only after delayed and often partial execution. This perspective, ScoAI = SC + OR + AI, argues that learned components should be designed around the decision they support. Supply-chain context defines the problem, operations research supplies objectives, constraints, and coordination, and AI supplies estimates and judgments that can be reused across tasks. We call a model or agent OR-native when these elements are explicit in its interface, and we refer to the coupling of learned components with executable operational relations as data--mechanism fusion. The coupling can fail in three places: when a judgment is turned into an action, when a locally good choice changes the cost of later stages, and when the executed action differs from the approved one. For each, we give a short analysis. Cost-sensitive thresholds and a decomposition of decision loss address the first; completion values and resource prices address the second; and a censoring argument shows how executed actions limit what operational logs can teach. Systems developed at JD.com illustrate these interfaces, and small constructed examples make every calculation reproducible. We close with decision records, an agent architecture, and evaluation protocols that compare reusable components with specialist workflows on decision quality, validity, coverage, and total cost.
- 10 October 2026
- Published by MDPI AG in Preprints.org
Open AccessScoAI: Toward OR-Native Decision Intelligence
Foundation models and AI agents are increasingly used to read operational requests, estimate uncertain quantities, and draft plans. In supply chains, these outputs feed decisions that must respect costs, stock and capacity limits, and business rules, and whose effects become visible only after delayed and often partial execution. This perspective, ScoAI = SC + OR + AI, argues that learned components should be designed around the decision they support. Supply-chain context defines the problem, operations research supplies objectives, constraints, and coordination, and AI supplies estimates and judgments that can be reused across tasks. We call a model or agent OR-native when these elements are explicit in its interface, and we refer to the coupling of learned components with executable operational relations as data--mechanism fusion. The coupling can fail in three places: when a judgment is turned into an action, when a locally good choice changes the cost of later stages, and when the executed action differs from the approved one. For each, we give a short analysis. Cost-sensitive thresholds and a decomposition of decision loss address the first; completion values and resource prices address the second; and a censoring argument shows how executed actions limit what operational logs can teach. Systems developed at JD.com illustrate these interfaces, and small constructed examples make every calculation reproducible. We close with decision records, an agent architecture, and evaluation protocols that compare reusable components with specialist workflows on decision quality, validity, coverage, and total cost.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11504-11515
Beyond Pairwise Interactions: Equivariant Hypergraph Diffusion for Crystal Structure Prediction
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 11504-11515
Open AccessBeyond Pairwise Interactions: Equivariant Hypergraph Diffusion for Crystal Structure Prediction
- 14 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. PP, 1-12
Meta-Heuristics Graph Neural Architecture Search for Combinatorial Optimization
Graph neural networks (GNNs) have emerged as a powerful tool for tackling Nondeterministic Polynomial-time (NP) hard combinatorial optimization (CO) problems which were believed that no algorithm can solve them in polynomial time, such as maximum cut and maximum independent set. The fundamental idea behind these methods is to represent CO problems as graphs, enabling GNNs to learn node or graph embeddings that encode essential combinatorial information. However, the design of GNN architectures for specific CO problems still demands considerable manual effort, relying heavily on domain expertise. Moreover, existing automated GNN architectures are primarily developed for traditional graph learning tasks and are not well-suited for addressing the unique challenges posed by NP-hard CO problems. To bridge this gap, we propose AutoGNP, a novel Automated GNN architecture search framework specifically designed to solve NP-hard combinatorial optimization problems. AutoGNP formulates CO problems using GNNs and focuses on two representative problem types: mixed integer linear programming and quadratic unconstrained binary optimization. The core idea of AutoGNP is to leverage graph neural architecture search algorithms to automatically identify the optimal GNN architecture for a given CO problem, significantly reducing the need for manual intervention and expertise. Unlike existing graph neural architecture search algorithms, AutoGNP incorporates 2-hop message passing operators in its architecture search space, allowing it to capture more intricate dependencies within the graph structure. We implement AutoGNP as an end-to-end learning framework, which is evaluated on benchmark combinatorial optimization datasets to demonstrate its performance. Our experimental results show that AutoGNP outperforms existing state-of-the-art GNN architectures across a variety of combinatorial optimization problems, yielding superior empirical performance.
- 14 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Emerging Topics in Computational Intelligence
- Vol. PP, 1-12
Meta-Heuristics Graph Neural Architecture Search for Combinatorial Optimization
Graph neural networks (GNNs) have emerged as a powerful tool for tackling Nondeterministic Polynomial-time (NP) hard combinatorial optimization (CO) problems which were believed that no algorithm can solve them in polynomial time, such as maximum cut and maximum independent set. The fundamental idea behind these methods is to represent CO problems as graphs, enabling GNNs to learn node or graph embeddings that encode essential combinatorial information. However, the design of GNN architectures for specific CO problems still demands considerable manual effort, relying heavily on domain expertise. Moreover, existing automated GNN architectures are primarily developed for traditional graph learning tasks and are not well-suited for addressing the unique challenges posed by NP-hard CO problems. To bridge this gap, we propose AutoGNP, a novel Automated GNN architecture search framework specifically designed to solve NP-hard combinatorial optimization problems. AutoGNP formulates CO problems using GNNs and focuses on two representative problem types: mixed integer linear programming and quadratic unconstrained binary optimization. The core idea of AutoGNP is to leverage graph neural architecture search algorithms to automatically identify the optimal GNN architecture for a given CO problem, significantly reducing the need for manual intervention and expertise. Unlike existing graph neural architecture search algorithms, AutoGNP incorporates 2-hop message passing operators in its architecture search space, allowing it to capture more intricate dependencies within the graph structure. We implement AutoGNP as an end-to-end learning framework, which is evaluated on benchmark combinatorial optimization datasets to demonstrate its performance. Our experimental results show that AutoGNP outperforms existing state-of-the-art GNN architectures across a variety of combinatorial optimization problems, yielding superior empirical performance.
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 3261-3272
FairCDR: Transferring Fairness and User Preferences for Cross-Domain Recommendation
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 3261-3272
Open AccessFairCDR: Transferring Fairness and User Preferences for Cross-Domain Recommendation
- Top Cited PapersOpen Access
- 15 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 43 (3) , 1-31
Contrastive Modality-Disentangled Learning for Multimodal Recommendation
Multimodal recommendation, which utilizes rich multimodal information to learn user preferences, has attracted significant attention. Most works focus on designing powerful encoders for extracting multimodal features, and simply aggregate the learned features together to make prediction. Consequently, they have a limited capacity to learn the inter-modality knowledge including the modality-shared and modality-unique knowledge. In fact, learning the modality-shared knowledge enables us to align cross-modality data for fusing heterogeneous modality features. Learning the modality-unique knowledge is equally important when recommendation tasks only involve a small amount of shared features and the necessary information is contained within specific modality. In this article, we propose Contrastive Modality-Disentangled Learning (CMDL) to overcome this critical limitation. CMDL exactly captures the inter-modality knowledge by achieving modality disentanglement. Specifically, CMDL first disentangles the initial representation into the modality-invariant and modality-specific representations. Afterwards, CMDL introduces a novel manner of contrastive learning to approximate the MI upper bounds for achieving disentanglement regularization. Building upon the proposed regularization, CMDL encourages the modality-invariant and modality-specific representations to capture the modality-shared and modality-unique knowledge respectively and to be statistically independent to each other. Empirically, extensive experiments are conducted on benchmark datasets, demonstrating the superior performance of CMDL compared with strong multimodal recommenders.
- 15 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Information Systems
- Vol. 43 (3) , 1-31
Top Cited PapersOpen AccessContrastive Modality-Disentangled Learning for Multimodal Recommendation
Multimodal recommendation, which utilizes rich multimodal information to learn user preferences, has attracted significant attention. Most works focus on designing powerful encoders for extracting multimodal features, and simply aggregate the learned features together to make prediction. Consequently, they have a limited capacity to learn the inter-modality knowledge including the modality-shared and modality-unique knowledge. In fact, learning the modality-shared knowledge enables us to align cross-modality data for fusing heterogeneous modality features. Learning the modality-unique knowledge is equally important when recommendation tasks only involve a small amount of shared features and the necessary information is contained within specific modality. In this article, we propose Contrastive Modality-Disentangled Learning (CMDL) to overcome this critical limitation. CMDL exactly captures the inter-modality knowledge by achieving modality disentanglement. Specifically, CMDL first disentangles the initial representation into the modality-invariant and modality-specific representations. Afterwards, CMDL introduces a novel manner of contrastive learning to approximate the MI upper bounds for achieving disentanglement regularization. Building upon the proposed regularization, CMDL encourages the modality-invariant and modality-specific representations to capture the modality-shared and modality-unique knowledge respectively and to be statistically independent to each other. Empirically, extensive experiments are conducted on benchmark datasets, demonstrating the superior performance of CMDL compared with strong multimodal recommenders.
- Open Access
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 2423-2432
MLN4KB: an efficient Markov logic network engine for large-scale knowledge bases and structured logic rules
Markov logic network (MLN) is a powerful statistical modeling framework for probabilistic logic reasoning. Despite the elegancy and effectiveness of MLN, the inference of MLN is known to suffer from an efficiency issue. Even the state-of-the-art MLN engines can not scale to medium-size real-world knowledge bases in the open-world setting, i.e., all unobserved facts in the knowledge base need predictions. In this work, by focusing on a certain class of first-order logic rules that are sufficiently expressive, we develop a highly efficient MLN inference engine called MLN4KB that can leverage the sparsity of knowledge bases. MLN4KB enjoys quite strong theoretical properties; its space and time complexities can be exponentially smaller than existing MLN engines. Experiments on both synthetic and real-world knowledge bases demonstrate the effectiveness of the proposed method. MLN4KB is orders of magnitudes faster (more than 103 times faster on some datasets) than existing MLN engines in the open-world setting. Without any approximation tricks, MLN4KB can scale to real-world knowledge bases including WN-18 and YAGO3-10 and achieve decent prediction accuracy without bells and whistles. We implement MLN4KB as a Julia package called MLN4KB.jl. The package supports both maximum a posteriori (MAP) inference and learning the weights of rules. MLN4KB.jl is public available at https://github.com/baidu-research/MLN4KB .
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 2423-2432
Open AccessMLN4KB: an efficient Markov logic network engine for large-scale knowledge bases and structured logic rules
Markov logic network (MLN) is a powerful statistical modeling framework for probabilistic logic reasoning. Despite the elegancy and effectiveness of MLN, the inference of MLN is known to suffer from an efficiency issue. Even the state-of-the-art MLN engines can not scale to medium-size real-world knowledge bases in the open-world setting, i.e., all unobserved facts in the knowledge base need predictions. In this work, by focusing on a certain class of first-order logic rules that are sufficiently expressive, we develop a highly efficient MLN inference engine called MLN4KB that can leverage the sparsity of knowledge bases. MLN4KB enjoys quite strong theoretical properties; its space and time complexities can be exponentially smaller than existing MLN engines. Experiments on both synthetic and real-world knowledge bases demonstrate the effectiveness of the proposed method. MLN4KB is orders of magnitudes faster (more than 103 times faster on some datasets) than existing MLN engines in the open-world setting. Without any approximation tricks, MLN4KB can scale to real-world knowledge bases including WN-18 and YAGO3-10 and achieve decent prediction accuracy without bells and whistles. We implement MLN4KB as a Julia package called MLN4KB.jl. The package supports both maximum a posteriori (MAP) inference and learning the weights of rules. MLN4KB.jl is public available at https://github.com/baidu-research/MLN4KB .
- Open Access
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 221-230
CurvDrop: A Ricci Curvature Based Approach to Prevent Graph Neural Networks from Over-Smoothing and Over-Squashing
Graph neural networks (GNNs) are powerful models to handle graph data and can achieve state-of-the-art in many critical tasks including node classification and link prediction. However, existing graph neural networks still face both challenges of over-smoothing and over-squashing based on previous literature. To this end, we propose a new Curvature-based topology-aware Dropout sampling technique named CurvDrop, in which we integrate the Discrete Ricci Curvature into graph neural networks to enable more expressive graph models. Also, this work can improve graph neural networks by quantifying connections in graphs and using structural information such as community structures in graphs. As a result, our method can tackle the both challenges of over-smoothing and over-squashing with theoretical justification. Also, numerous experiments on public datasets show the effectiveness and robustness of our proposed method. The code and data are released in https://github.com/liu-yang-maker/Curvature-based-Dropout.
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 221-230
Open AccessCurvDrop: A Ricci Curvature Based Approach to Prevent Graph Neural Networks from Over-Smoothing and Over-Squashing
Graph neural networks (GNNs) are powerful models to handle graph data and can achieve state-of-the-art in many critical tasks including node classification and link prediction. However, existing graph neural networks still face both challenges of over-smoothing and over-squashing based on previous literature. To this end, we propose a new Curvature-based topology-aware Dropout sampling technique named CurvDrop, in which we integrate the Discrete Ricci Curvature into graph neural networks to enable more expressive graph models. Also, this work can improve graph neural networks by quantifying connections in graphs and using structural information such as community structures in graphs. As a result, our method can tackle the both challenges of over-smoothing and over-squashing with theoretical justification. Also, numerous experiments on public datasets show the effectiveness and robustness of our proposed method. The code and data are released in https://github.com/liu-yang-maker/Curvature-based-Dropout.