Quanming Yao

Jianing Zhu, Jiangchao Yao, Tongliang Liu et al. · tsinghua

Privacy and security concerns in real-world applications have led to the development of adversarially robust federated models. However, the straightforward combination between adversarial training and federated learning in one framework can lead to the undesired robustness deterioration. We discover that the attribution behind this phenomenon is that the generated adversarial data could exacerbate the data heterogeneity among local clients, making the wrapped federated learning perform poorly. To deal with this problem, we propose a novel framework called Slack Federated Adversarial Training (SFAT), assigning the client-wise slack during aggregation to combat the intensified heterogeneity. Theoretically, we analyze the convergence of the proposed method to properly relax the objective when combining federated learning and adversarial training. Experimentally, we verify the rationality and effectiveness of SFAT on various benchmarked and real-world datasets with different adversarial training and federated optimization methods. The code is publicly available at https://github.com/ZFancy/SFAT.

Zhanke Zhou, Chenyu Zhou, Xuan Li et al. · tsinghua

Although powerful graph neural networks (GNNs) have boosted numerous real-world applications, the potential privacy risk is still underexplored. To close this gap, we perform the first comprehensive study of graph reconstruction attack that aims to reconstruct the adjacency of nodes. We show that a range of factors in GNNs can lead to the surprising leakage of private links. Especially by taking GNNs as a Markov chain and attacking GNNs via a flexible chain approximation, we systematically explore the underneath principles of graph reconstruction attack, and propose two information theory-guided mechanisms: (1) the chain-based attack method with adaptive designs for extracting more private information; (2) the chain-based defense method that sharply reduces the attack fidelity with moderate accuracy loss. Such two objectives disclose a critical belief that to recover better in attack, you must extract more multi-aspect knowledge from the trained GNN; while to learn safer for defense, you must forget more link-sensitive information in training GNNs. Empirically, we achieve state-of-the-art results on six datasets and three common GNNs. The code is publicly available at: https://github.com/tmlr-group/MC-GRA.

Zhen Wang, Haotong Du, Quanming Yao et al. · tsinghua

Completing missing facts is a fundamental task for temporal knowledge graphs (TKGs). Recently, graph neural network (GNN) based methods, which can simultaneously explore topological and temporal information, have become the state-of-the-art (SOTA) to complete TKGs. However, these studies are based on hand-designed architectures and fail to explore the diverse topological and temporal properties of TKG. To address this issue, we propose to use neural architecture search (NAS) to design data-specific message passing architecture for TKG completion. In particular, we develop a generalized framework to explore topological and temporal information in TKGs. Based on this framework, we design an expressive search space to fully capture various properties of different TKGs. Meanwhile, we adopt a search algorithm, which trains a supernet structure by sampling single path for efficient search with less cost. We further conduct extensive experiments on three benchmark datasets. The results show that the searched architectures by our method achieve the SOTA performances. Besides, the searched models can also implicitly reveal diverse properties in different TKGs. Our code is released in https://github.com/striderdu/SPA.

Lanning Wei, Huan Zhao, Xiaohan Zheng et al. · tsinghua

Designing versatile graph learning approaches is important, considering the diverse graphs and tasks existing in real-world applications. Existing methods have attempted to achieve this target through automated machine learning techniques, pre-training and fine-tuning strategies, and large language models. However, these methods are not versatile enough for graph learning, as they work on either limited types of graphs or a single task. In this paper, we propose to explore versatile graph learning approaches with LLM-based agents, and the key insight is customizing the graph learning procedures for diverse graphs and tasks. To achieve this, we develop several LLM-based agents, equipped with diverse profiles, tools, functions and human experience. They collaborate to configure each procedure with task and data-specific settings step by step towards versatile solutions, and the proposed method is dubbed GL-Agent. By evaluating on diverse tasks and graphs, the correct results of the agent and its comparable performance showcase the versatility of the proposed method, especially in complex scenarios.The low resource cost and the potential to use open-source LLMs highlight the efficiency of GL-Agent.

Zhen Xu, Lanning Wei, Huan Zhao et al. · tsinghua

Graph structured data is ubiquitous in daily life and scientific areas and has attracted increasing attention. Graph Neural Networks (GNNs) have been proved to be effective in modeling graph structured data and many variants of GNN architectures have been proposed. However, much human effort is often needed to tune the architecture depending on different datasets. Researchers naturally adopt Automated Machine Learning on Graph Learning, aiming to reduce the human effort and achieve generally top-performing GNNs, but their methods focus more on the architecture search. To understand GNN practitioners' automated solutions, we organized AutoGraph Challenge at KDD Cup 2020, emphasizing on automated graph neural networks for node classification. We received top solutions especially from industrial tech companies like Meituan, Alibaba and Twitter, which are already open sourced on Github. After detailed comparisons with solutions from academia, we quantify the gaps between academia and industry on modeling scope, effectiveness and efficiency, and show that (1) academia AutoML for Graph solutions focus on GNN architecture search while industrial solutions, especially the winning ones in the KDD Cup, tend to obtain an overall solution (2) by neural architecture search only, academia solutions achieve on average 97.3% accuracy of industrial solutions (3) academia solutions are cheap to obtain with several GPU hours while industrial solutions take a few months' labors. Academic solutions also contain much fewer parameters.

Yongqi Zhang, Zhanke Zhou, Quanming Yao et al.

Due to the popularity of Graph Neural Networks (GNNs), various GNN-based methods have been designed to reason on knowledge graphs (KGs). An important design component of GNN-based KG reasoning methods is called the propagation path, which contains a set of involved entities in each propagation step. Existing methods use hand-designed propagation paths, ignoring the correlation between the entities and the query relation. In addition, the number of involved entities will explosively grow at larger propagation steps. In this work, we are motivated to learn an adaptive propagation path in order to filter out irrelevant entities while preserving promising targets. First, we design an incremental sampling mechanism where the nearby targets and layer-wise connections can be preserved with linear complexity. Second, we design a learning-based sampling distribution to identify the semantically related entities. Extensive experiments show that our method is powerful, efficient, and semantic-aware. The code is available at https://github.com/LARS-research/AdaProp.

Yaqing Wang, Zaifei Yang, Quanming Yao · baidu

Background: Discovering potential drug-drug interactions (DDIs) is a long-standing challenge in clinical treatments and drug developments. Recently, deep learning techniques have been developed for DDI prediction. However, they generally require a huge number of samples, while known DDIs are rare. Methods: In this work, we present KnowDDI, a graph neural network-based method that addresses the above challenge. KnowDDI enhances drug representations by adaptively leveraging rich neighborhood information from large biomedical knowledge graphs. Then, it learns a knowledge subgraph for each drug-pair to interpret the predicted DDI, where each of the edges is associated with a connection strength indicating the importance of a known DDI or resembling strength between a drug-pair whose connection is unknown. Thus, the lack of DDIs is implicitly compensated by the enriched drug representations and propagated drug similarities. Results: We evaluate KnowDDI on two benchmark DDI datasets. Results show that KnowDDI obtains the state-of-the-art prediction performance with better interpretability. We also find that KnowDDI suffers less than existing works given a sparser knowledge graph. This indicates that the propagated drug similarities play a more important role in compensating for the lack of DDIs when the drug representations are less enriched. Conclusions: KnowDDI nicely combines the efficiency of deep learning techniques and the rich prior knowledge in biomedical knowledge graphs. As an original open-source tool, KnowDDI can help detect possible interactions in a broad range of relevant interaction prediction tasks, such as protein-protein interactions, drug-target interactions and disease-gene interactions, eventually promoting the development of biomedicine and healthcare.

Quanming Yao, James T. Kwok · tsinghua

Matrix and tensor completion aim to recover a low-rank matrix / tensor from limited observations and have been commonly used in applications such as recommender systems and multi-relational data mining. A state-of-the-art matrix completion algorithm is Soft-Impute, which exploits the special "sparse plus low-rank" structure of the matrix iterates to allow efficient SVD in each iteration. Though Soft-Impute is a proximal algorithm, it is generally believed that acceleration destroys the special structure and is thus not useful. In this paper, we show that Soft-Impute can indeed be accelerated without comprising this structure. To further reduce the iteration time complexity, we propose an approximate singular value thresholding scheme based on the power method. Theoretical analysis shows that the proposed algorithm still enjoys the fast $O(1/T^2)$ convergence rate of accelerated proximal algorithms. We further extend the proposed algorithm to tensor completion with the scaled latent nuclear norm regularizer. We show that a similar "sparse plus low-rank" structure also exists, leading to low iteration complexity and fast $O(1/T^2)$ convergence rate. Extensive experiments demonstrate that the proposed algorithm is much faster than Soft-Impute and other state-of-the-art matrix and tensor completion algorithms.

Ling Yue, Yongqi Zhang, Quanming Yao et al. · tencent-ai

Knowledge graph (KG) embedding is a fundamental task in natural language processing, and various methods have been proposed to explore semantic patterns in distinctive ways. In this paper, we propose to learn an ensemble by leveraging existing methods in a relation-aware manner. However, exploring these semantics using relation-aware ensemble leads to a much larger search space than general ensemble methods. To address this issue, we propose a divide-search-combine algorithm RelEns-DSC that searches the relation-wise ensemble weights independently. This algorithm has the same computation cost as general ensemble methods but with much better performance. Experimental results on benchmark datasets demonstrate the effectiveness of the proposed method in efficiently searching relation-aware ensemble weights and achieving state-of-the-art embedding performance. The code is public at https://github.com/LARS-research/RelEns.

Hui Zhang, Quanming Yao, James T. Kwok et al. · tsinghua

Feature extractor plays a critical role in text recognition (TR), but customizing its architecture is relatively less explored due to expensive manual tweaking. In this work, inspired by the success of neural architecture search (NAS), we propose to search for suitable feature extractors. We design a domain-specific search space by exploring principles for having good feature extractors. The space includes a 3D-structured space for the spatial model and a transformed-based space for the sequential model. As the space is huge and complexly structured, no existing NAS algorithms can be applied. We propose a two-stage algorithm to effectively search in the space. In the first stage, we cut the space into several blocks and progressively train each block with the help of an auxiliary head. We introduce the latency constraint into the second stage and search sub-network from the trained supernet via natural gradient descent. In experiments, a series of ablation studies are performed to better understand the designed space, search algorithm, and searched architectures. We also compare the proposed method with various state-of-the-art ones on both hand-written and scene TR tasks. Extensive results show that our approach can achieve better recognition performance with less latency.

Yinfeng Li, Chen Gao, Quanming Yao et al. · tsinghua

Spatiotemporal activity prediction, aiming to predict user activities at a specific location and time, is crucial for applications like urban planning and mobile advertising. Existing solutions based on tensor decomposition or graph embedding suffer from the following two major limitations: 1) ignoring the fine-grained similarities of user preferences; 2) user's modeling is entangled. In this work, we propose a hypergraph neural network model called DisenHCN to bridge the above gaps. In particular, we first unify the fine-grained user similarity and the complex matching between user preferences and spatiotemporal activity into a heterogeneous hypergraph. We then disentangle the user representations into different aspects (location-aware, time-aware, and activity-aware) and aggregate corresponding aspect's features on the constructed hypergraph, capturing high-order relations from different aspects and disentangles the impact of each aspect for final prediction. Extensive experiments show that our DisenHCN outperforms the state-of-the-art methods by 14.23% to 18.10% on four real-world datasets. Further studies also convincingly verify the rationality of each component in our DisenHCN.

Shiguang Wu, Yaqing Wang, Qinghe Jing et al. · baidu

Making personalized recommendation for cold-start users, who only have a few interaction histories, is a challenging problem in recommendation systems. Recent works leverage hypernetworks to directly map user interaction histories to user-specific parameters, which are then used to modulate predictor by feature-wise linear modulation function. These works obtain the state-of-the-art performance. However, the physical meaning of scaling and shifting in recommendation data is unclear. Instead of using a fixed modulation function and deciding modulation position by expertise, we propose a modulation framework called ColdNAS for user cold-start problem, where we look for proper modulation structure, including function and position, via neural architecture search. We design a search space which covers broad models and theoretically prove that this search space can be transformed to a much smaller space, enabling an efficient and robust one-shot search algorithm. Extensive experimental results on benchmark datasets show that ColdNAS consistently performs the best. We observe that different modulation functions lead to the best performance on different datasets, which validates the necessity of designing a searching-based method.

Lanning Wei, Zhiqiang He, Huan Zhao et al. · tsinghua

In recent years, Graph Neural Networks (GNNs) have been popular in the graph classification task. Currently, shallow GNNs are more common due to the well-known over-smoothing problem facing deeper GNNs. However, they are sub-optimal without utilizing the information from distant nodes, i.e., the long-range dependencies. The mainstream methods in the graph classification task can extract the long-range dependencies either by designing the pooling operations or incorporating the higher-order neighbors, while they have evident drawbacks by modifying the original graph structure, which may result in information loss in graph structure learning. In this paper, by justifying the smaller influence of the over-smoothing problem in the graph classification task, we evoke the importance of stacking-based GNNs and then employ them to capture the long-range dependencies without modifying the original graph structure. To achieve this, two design needs are given for stacking-based GNNs, i.e., sufficient model depth and adaptive skip-connection schemes. By transforming the two design needs into designing data-specific inter-layer connections, we propose a novel approach with the help of neural architecture search (NAS), which is dubbed LRGNN (Long-Range Graph Neural Networks). Extensive experiments on five datasets show that the proposed LRGNN can achieve the best performance, and obtained data-specific GNNs with different depth and skip-connection schemes, which can better capture the long-range dependencies.

Quanming Yao, Yaqing Wang, Bo Han et al. · baidu, tsinghua

Nonconvex regularization has been popularly used in low-rank matrix learning. However, extending it for low-rank tensor learning is still computationally expensive. To address this problem, we develop an efficient solver for use with a nonconvex extension of the overlapped nuclear norm regularizer. Based on the proximal average algorithm, the proposed algorithm can avoid expensive tensor folding/unfolding operations. A special "sparse plus low-rank" structure is maintained throughout the iterations, and allows fast computation of the individual proximal steps. Empirical convergence is further improved with the use of adaptive momentum. We provide convergence guarantees to critical points on smooth losses and also on objectives satisfying the Kurdyka-Łojasiewicz condition. While the optimization problem is nonconvex and nonsmooth, we show that its critical points still have good statistical performance on the tensor completion problem. Experiments on various synthetic and real-world data sets show that the proposed algorithm is efficient in both time and space and more accurate than the existing state-of-the-art.

Zhanke Zhou, Jiangchao Yao, Jiaxu Liu et al.

Although link prediction on graphs has achieved great success with the development of graph neural networks (GNNs), the potential robustness under the edge noise is still less investigated. To close this gap, we first conduct an empirical study to disclose that the edge noise bilaterally perturbs both input topology and target label, yielding severe performance degradation and representation collapse. To address this dilemma, we propose an information-theory-guided principle, Robust Graph Information Bottleneck (RGIB), to extract reliable supervision signals and avoid representation collapse. Different from the basic information bottleneck, RGIB further decouples and balances the mutual dependence among graph topology, target labels, and representation, building new learning objectives for robust representation against the bilateral noise. Two instantiations, RGIB-SSL and RGIB-REP, are explored to leverage the merits of different methodologies, i.e., self-supervised learning and data reparameterization, for implicit and explicit data denoising, respectively. Extensive experiments on six datasets and three GNNs with diverse noisy scenarios verify the effectiveness of our RGIB instantiations. The code is publicly available at: https://github.com/tmlr-group/RGIB.

Yongqi Zhang, Zhanke Zhou, Quanming Yao et al. · tsinghua

While hyper-parameters (HPs) are important for knowledge graph (KG) learning, existing methods fail to search them efficiently. To solve this problem, we first analyze the properties of different HPs and measure the transfer ability from small subgraph to the full graph. Based on the analysis, we propose an efficient two-stage search algorithm KGTuner, which efficiently explores HP configurations on small subgraph at the first stage and transfers the top-performed configurations for fine-tuning on the large full graph at the second stage. Experiments show that our method can consistently find better HPs than the baseline algorithms within the same time budget, which achieves {9.1\%} average relative improvement for four embedding models on the large-scale KGs in open graph benchmark.

Yongqi Zhang, Quanming Yao, Ling Yue et al. · tencent-ai

Accurately predicting drug-drug interactions (DDI) for emerging drugs, which offer possibilities for treating and alleviating diseases, with computational methods can improve patient care and contribute to efficient drug development. However, many existing computational methods require large amounts of known DDI information, which is scarce for emerging drugs. In this paper, we propose EmerGNN, a graph neural network (GNN) that can effectively predict interactions for emerging drugs by leveraging the rich information in biomedical networks. EmerGNN learns pairwise representations of drugs by extracting the paths between drug pairs, propagating information from one drug to the other, and incorporating the relevant biomedical concepts on the paths. The different edges on the biomedical network are weighted to indicate the relevance for the target DDI prediction. Overall, EmerGNN has higher accuracy than existing approaches in predicting interactions for emerging drugs and can identify the most relevant information on the biomedical network.

Haiquan Qiu, Yongqi Zhang, Yong Li et al.

Rule learning is critical to improving knowledge graph (KG) reasoning due to their ability to provide logical and interpretable explanations. Recently, Graph Neural Networks (GNNs) with tail entity scoring achieve the state-of-the-art performance on KG reasoning. However, the theoretical understandings for these GNNs are either lacking or focusing on single-relational graphs, leaving what the kind of rules these GNNs can learn an open problem. We propose to fill the above gap in this paper. Specifically, GNNs with tail entity scoring are unified into a common framework. Then, we analyze their expressivity by formally describing the rule structures they can learn and theoretically demonstrating their superiority. These results further inspire us to propose a novel labeling strategy to learn more rules in KG reasoning. Experimental results are consistent with our theoretical findings and verify the effectiveness of our proposed method. The code is publicly available at https://github.com/LARS-research/Rule-learning-expressivity.

Xu Wang, Huan Zhao, Lanning Wei et al. · tsinghua

Aiming at two molecular graph datasets and one protein association subgraph dataset in OGB graph classification task, we design a graph neural network framework for graph classification task by introducing PAS(Pooling Architecture Search). At the same time, we improve it based on the GNN topology design method F2GNN to further design the feature selection and fusion strategies, so as to further improve the performance of the model in the graph property prediction task while overcoming the over smoothing problem of deep GNN training. Finally, a performance breakthrough is achieved on these three datasets, which is significantly better than other methods with fixed aggregate function. It is proved that the NAS method has high generalization ability for multiple tasks and the advantage of our method in processing graph property prediction tasks.

Lanning Wei, Zhiqiang He, Huan Zhao et al. · tsinghua

In recent years, Graph Neural Networks (GNNs) have been popular in graph representation learning which assumes the homophily property, i.e., the connected nodes have the same label or have similar features. However, they may fail to generalize into the heterophilous graphs which in the low/medium level of homophily. Existing methods tend to address this problem by enhancing the intra-class information extraction, i.e., either by designing better GNNs to improve the model effectiveness, or re-designing the graph structures to incorporate more potential intra-class nodes from distant hops. Despite the success, we observe two aspects that can be further improved: (a) enhancing the ego feature information extraction from node itself which is more reliable in extracting the intra-class information; (b) designing node-wise GNNs can better adapt to the nodes with different homophily ratios. In this paper, we propose a novel method IIE-GNN (Intra-class Information Enhanced Graph Neural Networks) to achieve two improvements. A unified framework is proposed based on the literature, in which the intra-class information from the node itself and neighbors can be extracted based on seven carefully designed blocks. With the help of neural architecture search (NAS), we propose a novel search space based on the framework, and then provide an architecture predictor to design GNNs for each node. We further conduct experiments to show that IIE-GNN can improve the model performance by designing node-wise GNNs to enhance intra-class information extraction.

Hansi Yang, Yongqi Zhang, Quanming Yao · tsinghua

Scoring function (SF) measures the plausibility of triplets in knowledge graphs. Different scoring functions can lead to huge differences in link prediction performances on different knowledge graphs. In this report, we describe a weird scoring function found by random search on the open graph benchmark (OGB). This scoring function, called AutoWeird, only uses tail entity and relation in a triplet to compute its plausibility score. Experimental results show that AutoWeird achieves top-1 performance on ogbl-wikikg2 data set, but has much worse performance than other methods on ogbl-biokg data set. By analyzing the tail entity distribution and evaluation protocol of these two data sets, we attribute the unexpected success of AutoWeird on ogbl-wikikg2 to inappropriate evaluation and concentrated tail entity distribution. Such results may motivate further research on how to accurately evaluate the performance of different link prediction methods for knowledge graphs.

Yihong Chen, Quanming Yao

Attention sinks and massive activations are recurring and closely related phenomena in Transformer models. Existing studies have largely focused on the forward pass, making it unclear whether their connection is direct or mediated by a training-time mechanism. We study this question from the perspective of backpropagation. Empirically and theoretically, we show that under causal mask, attention sinks can induce pronounced gradient concentration, which we term gradient sinks. Furthermore, in pre-norm architectures with RMSNorm, massive activations can be understood as an adaptive response to this localized gradient pressure during training. To test this hypothesis, we introduce V-scale, a modification that adjusts value-path backpropagated gradients. In pretrained V-scale models, attention sinks are preserved whereas massive activations are suppressed. These results support the interpretation that gradient sink is a key training-time mediator linking attention sinks and massive activations.

Xu Wang, Huan Zhao, Weiwei Tu et al.

Molecular property prediction is an important problem in drug discovery and materials science. As geometric structures have been demonstrated necessary for molecular property prediction, 3D information has been combined with various graph learning methods to boost prediction performance. However, obtaining the geometric structure of molecules is not feasible in many real-world applications due to the high computational cost. In this work, we propose a novel 3D pre-training framework (dubbed 3D PGT), which pre-trains a model on 3D molecular graphs, and then fine-tunes it on molecular graphs without 3D structures. Based on fact that bond length, bond angle, and dihedral angle are three basic geometric descriptors corresponding to a complete molecular 3D conformer, we first develop a multi-task generative pre-train framework based on these three attributes. Next, to automatically fuse these three generative tasks, we design a surrogate metric using the \textit{total energy} to search for weight distribution of the three pretext task since total energy corresponding to the quality of 3D conformer.Extensive experiments on 2D molecular graphs are conducted to demonstrate the accuracy, efficiency and generalization ability of the proposed 3D PGT compared to various pre-training baselines.

Hansi Yang, Yongqi Zhang, Quanming Yao et al.

Node classification on graphs is an important research problem with many applications. Real-world graph data sets may not be balanced and accurate as assumed by most existing works. A challenging setting is positive-unlabeled (PU) node classification, where labeled nodes are restricted to positive nodes. It has diverse applications, e.g., pandemic prediction or network anomaly detection. Existing works on PU node classification overlook information in the graph structure, which can be critical. In this paper, we propose to better utilize graph structure for PU node classification. We first propose a distance-aware PU loss that uses homophily in graphs to introduce more accurate supervision. We also propose a regularizer to align the model with graph structure. Theoretical analysis shows that minimizing the proposed loss also leads to minimizing the expected loss with both positive and negative labels. Extensive empirical evaluation on diverse graph data sets demonstrates its superior performance over existing state-of-the-art methods.

Zhen Hao Wong, Ling Yue, Quanming Yao

Graph Neural Networks (GNNs) have shown success in various fields for learning from graph-structured data. This paper investigates the application of ensemble learning techniques to improve the performance and robustness of Graph Neural Networks (GNNs). By training multiple GNN models with diverse initializations or architectures, we create an ensemble model named ELGNN that captures various aspects of the data and uses the Tree-Structured Parzen Estimator algorithm to determine the ensemble weights. Combining the predictions of these models enhances overall accuracy, reduces bias and variance, and mitigates the impact of noisy data. Our findings demonstrate the efficacy of ensemble learning in enhancing GNN capabilities for analyzing complex graph-structured data. The code is public at https://github.com/wongzhenhao/ELGNN.

Bingyu Hui, Lebin Yu, Quanming Yao et al.

Zero-shot coordination(ZSC), a key challenge in multi-agent game theory, has become a hot topic in reinforcement learning (RL) research recently, especially in complex evolving games. It focuses on the generalization ability of agents, requiring them to coordinate well with collaborators from a diverse, potentially evolving, pool of partners that are not seen before without any fine-tuning. Population-based training, which approximates such an evolving partner pool, has been proven to provide good zero-shot coordination performance; nevertheless, existing methods are limited by computational resources, mainly focusing on optimizing diversity in small populations while neglecting the potential performance gains from scaling population size. To address this issue, this paper proposes the Scalable Population Training (ScaPT), an efficient RL training framework comprising two key components: a meta-agent that efficiently realizes a population by selectively sharing parameters across agents, and a mutual information regularizer that guarantees population diversity. To empirically validate the effectiveness of ScaPT, this paper evaluates it along with representational frameworks in Hanabi cooperative game and confirms its superiority.

Haiquan Qiu, Quanming Yao

The pursuit of computational efficiency has driven the adoption of low-precision formats for training transformer models. However, this progress is often hindered by notorious training instabilities. This paper provides the first mechanistic explanation for a long-standing and unresolved failure case where training with flash attention in low-precision settings leads to catastrophic loss explosion. Our in-depth analysis reveals that the failure is not a random artifact but caused by two intertwined phenomena: the emergence of similar low-rank representations within the attention mechanism and the compounding effect of biased rounding errors inherent in low-precision arithmetic. We demonstrate how these factors create a vicious cycle of error accumulation that corrupts weight updates, ultimately derailing the training dynamics. To validate our findings, we introduce a minimal modification to the flash attention that mitigates the bias in rounding errors. This simple change stabilizes the training process, confirming our analysis and offering a practical solution to this persistent problem. Code is available at https://github.com/ucker/why-low-precision-training-fails.

Zhenqian Shen, Mingyang Zhou, Yongqi Zhang et al.

Motivation: Emerging drug-drug interaction (DDI) prediction is crucial for new drugs but is hindered by distribution changes between known and new drugs in real-world scenarios. Current evaluation often neglects these changes, relying on unrealistic i.i.d. split due to the absence of drug approval data. Results: We propose DDI-Ben, a benchmarking framework for emerging DDI prediction under distribution changes. DDI-Ben introduces a distribution change simulation framework that leverages distribution changes between drug sets as a surrogate for real-world distribution changes of DDIs, and is compatible with various drug split strategies. Through extensive benchmarking on ten representative methods, we show that most existing approaches suffer substantial performance degradation under distribution changes. Our analysis further indicates that large language model (LLM) based methods and the integration of drug-related textual information offer promising robustness against such degradation. To support future research, we release the benchmark datasets with simulated distribution changes. Overall, DDI-Ben highlights the importance of explicitly addressing distribution changes and provides a foundation for developing more resilient methods for emerging DDI prediction. Availability and implementation: Our code and data are available at https://github.com/LARS-research/DDI-Bench.

Zhanke Zhou, Yongqi Zhang, Jiangchao Yao et al.

To deduce new facts on a knowledge graph (KG), a link predictor learns from the graph structure and collects local evidence to find the answer to a given query. However, existing methods suffer from a severe scalability problem due to the utilization of the whole KG for prediction, which hinders their promise on large scale KGs and cannot be directly addressed by vanilla sampling methods. In this work, we propose the one-shot-subgraph link prediction to achieve efficient and adaptive prediction. The design principle is that, instead of directly acting on the whole KG, the prediction procedure is decoupled into two steps, i.e., (i) extracting only one subgraph according to the query and (ii) predicting on this single, query dependent subgraph. We reveal that the non-parametric and computation-efficient heuristics Personalized PageRank (PPR) can effectively identify the potential answers and supporting evidence. With efficient subgraph-based prediction, we further introduce the automated searching of the optimal configurations in both data and model spaces. Empirically, we achieve promoted efficiency and leading performances on five large-scale benchmarks. The code is publicly available at: https://github.com/tmlr-group/one-shot-subgraph.

Zhen Xu, Sergio Escalera, Isabelle Guyon et al.

Obtaining standardized crowdsourced benchmark of computational methods is a major issue in data science communities. Dedicated frameworks enabling fair benchmarking in a unified environment are yet to be developed. Here we introduce Codabench, an open-source, community-driven platform for benchmarking algorithms or software agents versus datasets or tasks. A public instance of Codabench (https://www.codabench.org/) is open to everyone, free of charge, and allows benchmark organizers to compare fairly submissions, under the same setting (software, hardware, data, algorithms), with custom protocols and data formats. Codabench has unique features facilitating the organization of benchmarks flexibly, easily and reproducibly, such as the possibility of re-using templates of benchmarks, and supplying compute resources on-demand. Codabench has been used internally and externally on various applications, receiving more than 130 users and 2500 submissions. As illustrative use cases, we introduce 4 diverse benchmarks covering Graph Machine Learning, Cancer Heterogeneity, Clinical Diagnosis and Reinforcement Learning.

Huan Zhao, Quanming Yao, Weiwei Tu

Recent years have witnessed the popularity and success of graph neural networks (GNN) in various scenarios. To obtain data-specific GNN architectures, researchers turn to neural architecture search (NAS), which has made impressive success in discovering effective architectures in convolutional neural networks. However, it is non-trivial to apply NAS approaches to GNN due to challenges in search space design and the expensive searching cost of existing NAS methods. In this work, to obtain the data-specific GNN architectures and address the computational challenges facing by NAS approaches, we propose a framework, which tries to Search to Aggregate NEighborhood (SANE), to automatically design data-specific GNN architectures. By designing a novel and expressive search space, we propose a differentiable search algorithm, which is more efficient than previous reinforcement learning based methods. Experimental results on four tasks and seven real-world datasets demonstrate the superiority of SANE compared to existing GNN models and NAS approaches in terms of effectiveness and efficiency. (Code is available at: https://github.com/AutoML-4Paradigm/SANE).

Fengli Xu, Quanming Yao, Pan Hui et al.

Distinguishing the automorphic equivalence of nodes in a graph plays an essential role in many scientific domains, e.g., computational biologist and social network analysis. However, existing graph neural networks (GNNs) fail to capture such an important property. To make GNN aware of automorphic equivalence, we first introduce a localized variant of this concept -- ego-centered automorphic equivalence (Ego-AE). Then, we design a novel variant of GNN, i.e., GRAPE, that uses learnable AE-aware aggregators to explicitly differentiate the Ego-AE of each node's neighbors with the aids of various subgraph templates. While the design of subgraph templates can be hard, we further propose a genetic algorithm to automatically search them from graph data. Moreover, we theoretically prove that GRAPE is expressive in terms of generating distinct representations for nodes with different Ego-AE features, which fills in a fundamental gap of existing GNN variants. Finally, we empirically validate our model on eight real-world graph data, including social network, e-commerce co-purchase network, and citation network, and show that it consistently outperforms existing GNNs. The source code is public available at https://github.com/tsinghua-fib-lab/GRAPE.

Ziying Zhang, Quanming Yao, Yaqing Wang

In this work, we tackle the challenge of recommending emerging items, whose interactions gradually accumulate over time. Existing methods often overlook this dynamic process, typically assuming that emerging items have few or even no historical interactions. Such an assumption oversimplifies the problem, as a good model must preserve the uniqueness of emerging items while leveraging their shared patterns with established ones. To address this challenge, we propose EmerFlow, a novel LLM-empowered representation learning framework that generates distinctive embeddings for emerging items. It first enriches the raw features of emerging items through LLM reasoning, then aligns these representations with the embedding space of the existing recommendation model. Finally, new interactions are incorporated through meta-learning to refine the embeddings. This enables EmerFlow to learn expressive embeddings for emerging items from only limited interactions. Extensive experiments across diverse domains, including movies and pharmaceuticals, show that EmerFlow consistently outperforms existing methods.

Guangyi Liu, Quanming Yao, Yongqi Zhang et al.

Recommendation systems, as widely implemented nowadays on various platforms, recommend relevant items to users based on their preferences. The classical methods which rely on user-item interaction matrices has limitations, especially in scenarios where there is a lack of interaction data for new items. Knowledge graph (KG)-based recommendation systems have emerged as a promising solution. However, most KG-based methods adopt node embeddings, which do not provide personalized recommendations for different users and cannot generalize well to the new items. To address these limitations, we propose Knowledge-enhanced User-Centric subgraph Network (KUCNet), a subgraph learning approach with graph neural network (GNN) for effective recommendation. KUCNet constructs a U-I subgraph for each user-item pair that captures both the historical information of user-item interactions and the side information provided in KG. An attention-based GNN is designed to encode the U-I subgraphs for recommendation. Considering efficiency, the pruned user-centric computation graph is further introduced such that multiple U-I subgraphs can be simultaneously computed and that the size can be pruned by Personalized PageRank. Our proposed method achieves accurate, efficient, and interpretable recommendations especially for new items. Experimental results demonstrate the superiority of KUCNet over state-of-the-art KG-based and collaborative filtering (CF)-based methods.

Lebin Yu, Yunbo Qiu, Quanming Yao et al.

Communication in multi-agent reinforcement learning (MARL) has been proven to effectively promote cooperation among agents recently. Since communication in real-world scenarios is vulnerable to noises and adversarial attacks, it is crucial to develop robust communicative MARL technique. However, existing research in this domain has predominantly focused on passive defense strategies, where agents receive all messages equally, making it hard to balance performance and robustness. We propose an active defense strategy, where agents automatically reduce the impact of potentially harmful messages on the final decision. There are two challenges to implement this strategy, that are defining unreliable messages and adjusting the unreliable messages' impact on the final decision properly. To address them, we design an Active Defense Multi-Agent Communication framework (ADMAC), which estimates the reliability of received messages and adjusts their impact on the final decision accordingly with the help of a decomposable decision structure. The superiority of ADMAC over existing methods is validated by experiments in three communication-critical tasks under four types of attacks.

Haotong Du, Yaqing Wang, Fei Xiong et al.

With the expansion of business scales and scopes on online platforms, multi-scenario matching has become a mainstream solution to reduce maintenance costs and alleviate data sparsity. The key to effective multi-scenario recommendation lies in capturing both user preferences shared across all scenarios and scenario-aware preferences specific to each scenario. However, existing methods often overlook user-specific modeling, limiting the generation of personalized user representations. To address this, we propose PERSCEN, an innovative approach that incorporates user-specific modeling into multi-scenario matching. PERSCEN constructs a user-specific feature graph based on user characteristics and employs a lightweight graph neural network to capture higher-order interaction patterns, enabling personalized extraction of preferences shared across scenarios. Additionally, we leverage vector quantization techniques to distil scenario-aware preferences from users' behavior sequence within individual scenarios, facilitating user-specific and scenario-aware preference modeling. To enhance efficient and flexible information transfer, we introduce a progressive scenario-aware gated linear unit that allows fine-grained, low-latency fusion. Extensive experiments demonstrate that PERSCEN outperforms existing methods. Further efficiency analysis confirms that PERSCEN effectively balances performance with computational cost, ensuring its practicality for real-world industrial systems.

Haiquan Qiu, Shuzhi Liu, Quanming Yao

Complex networks describe important structures in nature and society, composed of nodes and the edges that connect them. The evolution of these networks is typically described by dynamics, which are labor-intensive and require expert knowledge to derive. However, because the complex network involves noisy observations from multiple trajectories of nodes, existing symbolic regression methods are either not applicable or ineffective on its dynamics. In this paper, we propose Physically Inspired Neural Dynamics Symbolic Regression (PI-NDSR), a method based on neural networks and genetic programming to automatically learn the symbolic expression of dynamics. Our method consists of two key components: a Physically Inspired Neural Dynamics (PIND) to augment and denoise trajectories through observed trajectory interpolation; and a coordinated genetic search algorithm to derive symbolic expressions. This algorithm leverages references of node dynamics and edge dynamics from neural dynamics to avoid overfitted expressions in symbolic space. We evaluate our method on synthetic datasets generated by various dynamics and real datasets on disease spreading. The results demonstrate that PI-NDSR outperforms the existing method in terms of both recovery probability and error.

Haiquan Qiu, Yatao Bian, Quanming Yao

Message passing mechanism contributes to the success of GNNs in various applications, but also brings the oversquashing problem. Recent works combat oversquashing by improving the graph spectrums with rewiring techniques, disrupting the structural bias in graphs, and having limited improvement on oversquashing in terms of oversquashing measure. Motivated by unitary RNN, we propose Graph Unitary Message Passing (GUMP) to alleviate oversquashing in GNNs by applying unitary adjacency matrix for message passing. To design GUMP, a transformation is first proposed to make general graphs have unitary adjacency matrix and keep its structural bias. Then, unitary adjacency matrix is obtained with a unitary projection algorithm, which is implemented by utilizing the intrinsic structure of unitary adjacency matrix and allows GUMP to be permutation-equivariant. Experimental results show the effectiveness of GUMP in improving the performance on various graph learning tasks.

Ruibing Wang, Shuhan Guo, Zeen Li et al.

Traffic Signal Control (TSC) involves a challenging trade-off: classic heuristics are efficient but oversimplified, while Deep Reinforcement Learning (DRL) achieves high performance yet suffers from poor generalization and opaque policies. Online Large Language Models (LLMs) provide general reasoning but incur high latency and lack environment-specific optimization. To address these issues, we propose Temporal Policy Evolution for Traffic (\textbf{\method{}}), which uses LLMs as an evolution engine to derive specialized heuristic policies. The framework introduces two key modules: (1) Structured State Abstraction (SSA), converting high-dimensional traffic data into temporal-logical facts for reasoning; and (2) Credit Assignment Feedback (CAF), tracing flawed micro-decisions to poor macro-outcomes for targeted critique. Operating entirely at the prompt level without training, \method{} yields lightweight, robust policies optimized for specific traffic environments, outperforming both heuristics and online LLM actors.

Haiquan Qiu, You Wu, Yingjie Tan et al.

Loss explosions in training deep neural networks can nullify multi-million dollar training runs. Conventional monitoring metrics like weight and gradient norms are often lagging and ambiguous predictors, as their values vary dramatically across different models and even between layers of the same model, making it difficult to establish a unified standard for detecting impending failure. We introduce Spectral Alignment (SA), a novel, theoretically-grounded metric that monitors the distributional alignment between layer inputs and the principal singular vectors of weight matrices. We show that a collapse in the sign diversity of this alignment is a powerful early predictor of representational collapse and training divergence. Empirical results on language models demonstrate that monitoring the SA distribution provides a significantly earlier and clearer warning of loss explosions than traditional scalar metrics. SA's low computational overhead makes it a practical tool for safeguarding model training.

Ziying Zhang, Yaqing Wang, Yuxuan Sun et al.

Cold-start drug-target interaction (DTI) prediction focuses on interaction between novel drugs and proteins. Previous methods typically learn transferable interaction patterns between structures of drug and proteins to tackle it. However, insight from proteomics suggest that protein have multi-level structures and they all influence the DTI. Existing works usually represent protein with only primary structures, limiting their ability to capture interactions involving higher-level structures. Inspired by this insight, we propose ColdDTI, a framework attending on protein multi-level structure for cold-start DTI prediction. We employ hierarchical attention mechanism to mine interaction between multi-level protein structures (from primary to quaternary) and drug structures at both local and global granularities. Then, we leverage mined interactions to fuse structure representations of different levels for final prediction. Our design captures biologically transferable priors, avoiding the risk of overfitting caused by excessive reliance on representation learning. Experiments on benchmark datasets demonstrate that ColdDTI consistently outperforms previous methods in cold-start settings.

Ziying Zhang, Yaqing Wang, Quanming Yao

Meta reasoning behaviors work as a skeleton to guide large language model (LLM) reasoning, thus help to improve reasoning performance. However, prior researches implement meta reasoning skeleton with manually designed structure, limiting ability to adapt to query-specific requirement and capture intricate logical dependency among reasoning steps. To deal with the challenges, we represent meta reasoning skeleton with directed acyclic graph (DAG) to unify skeletons proposed in prior works and model intricate logical dependency. Then we propose AutoMR, a framework that searches for query-aware meta reasoning skeleton automatically inspired by automated machine learning (AutoML). Specifically, we construct search space based on DAG representation of skeleton and then formulate the search problem. We design a dynamic skeleton sampling algorithm by expanding meta reasoning skeleton along with reasoning context at inference time. This algorithm can derive any meta reasoning skeleton in search space efficiently and adapt skeleton to evolving base reasoning context, thus enable efficient query-aware skeleton search. We conduct experiments on extensive benchmark datasets. Experimental results show that AutoMR achieves better reasoning performance than previous works broadly.

Guangyi Liu, Yongqi Zhang, Xunyuan Liu et al.

Drug-drug interaction (DDI) prediction is critical for treatment safety. While large language models (LLMs) show promise in pharmaceutical tasks, their effectiveness in DDI prediction remains challenging. Inspired by the well-established clinical practice where physicians routinely reference similar historical cases to guide their decisions through case-based reasoning (CBR), we propose CBR-DDI, a novel framework that distills pharmacological principles from historical cases to improve LLM reasoning for DDI tasks. CBR-DDI constructs a knowledge repository by leveraging LLMs to extract pharmacological insights and graph neural networks (GNNs) to model drug associations. A hybrid retrieval mechanism and dual-layer knowledge-enhanced prompting allow LLMs to effectively retrieve and reuse relevant cases. We further introduce a representative sampling strategy for dynamic case refinement. Extensive experiments demonstrate that CBR-DDI achieves state-of-the-art performance, with a significant 28.7% accuracy improvement over both popular LLMs and CBR baseline, while maintaining high interpretability and flexibility.

Shiguang Wu, Yaqing Wang, Quanming Yao

As higher-level intelligence emerges from the combination of modular components with lower-level intelligence, many works combines Large Language Models (LLMs) for collective intelligence. Such combination is achieved by building communications among LLMs. While current systems primarily facilitate such communication through natural language, this paper proposes a novel paradigm of direct dense vector communication between LLMs. Our approach eliminates the unnecessary embedding and de-embedding steps when LLM interact with another, enabling more efficient information transfer, fully differentiable optimization pathways, and exploration of capabilities beyond human heuristics. We use such stripped LLMs as vertexes and optimizable seq2seq modules as edges to construct LMNet, with similar structure as MLPs. By utilizing smaller pre-trained LLMs as vertexes, we train a LMNet that achieves comparable performance with LLMs in similar size with only less than 0.1% training cost. This offers a new perspective on scaling for general intelligence rather than training a monolithic LLM from scratch. Besides, the proposed method can be used for other applications, like customizing LLM with limited data, showing its versatility.

Quanming Yao, Yongqi Zhang, Yaqing Wang et al.

The brute-force scaleup of training datasets, learnable parameters and computation power, has become a prevalent strategy for developing more robust learning models. However, due to bottlenecks in data, computation, and trust, the sustainability of this strategy is a serious concern. In this paper, we attempt to address this issue in a parsimonious manner (i.e., achieving greater potential with simpler models). The key is to drive models using domain-specific knowledge, such as symbols, logic, and formulas, instead of purely relying on scaleup. This approach allows us to build a framework that uses this knowledge as "building blocks" to achieve parsimony in model design, training, and interpretation. Empirical results show that our methods surpass those that typically follow the scaling law. We also demonstrate our framework in AI for science, specifically in the problem of drug-drug interaction prediction. We hope our research can foster more diverse technical roadmaps in the era of foundation models.

Guangyi Liu, Yongqi Zhang, Yong Li et al.

Large Language Models (LLMs) excel at intuitive, implicit reasoning. Guiding LLMs to construct thought chains can enhance their deliberate reasoning abilities, but also faces challenges such as hallucination. Knowledge Graphs (KGs) can provide explicit structured knowledge for LLMs to alleviate these issues. However, existing KG-enhanced methods often overlook explicit graph learning, making it challenging to efficiently provide precise reasoning chains for LLMs. Following dual-process theory, we propose Dual-Reasoning (DualR), a novel framework that integrates an external system based on Graph Neural Network (GNN) for explicit reasoning on KGs, complementing the implicit reasoning of LLMs through externalized reasoning chains. DualR designs an LLM-empowered GNN module for explicit learning on KGs, efficiently extracting high-quality reasoning chains. These reasoning chains are then refined to a knowledge-enhanced multiple-choice prompt, guiding a frozen LLM to reason thoughtfully for final answer determination. Extensive experiments on three benchmark KGQA datasets demonstrate that DualR achieves state-of-the-art performance while maintaining high efficiency and interpretability.

Huiming Chen, Huandong Wang, Quanming Yao et al.

Federated optimization (FedOpt), which targets at collaboratively training a learning model across a large number of distributed clients, is vital for federated learning. The primary concerns in FedOpt can be attributed to the model divergence and communication efficiency, which significantly affect the performance. In this paper, we propose a new method, i.e., LoSAC, to learn from heterogeneous distributed data more efficiently. Its key algorithmic insight is to locally update the estimate for the global full gradient after {each} regular local model update. Thus, LoSAC can keep clients' information refreshed in a more compact way. In particular, we have studied the convergence result for LoSAC. Besides, the bonus of LoSAC is the ability to defend the information leakage from the recent technique Deep Leakage Gradients (DLG). Finally, experiments have verified the superiority of LoSAC comparing with state-of-the-art FedOpt algorithms. Specifically, LoSAC significantly improves communication efficiency by more than $100\%$ on average, mitigates the model divergence problem and equips with the defense ability against DLG.

Yaqing Wang, Song Wang, Quanming Yao et al.

Short text classification is a fundamental task in natural language processing. It is hard due to the lack of context information and labeled data in practice. In this paper, we propose a new method called SHINE, which is based on graph neural network (GNN), for short text classification. First, we model the short text dataset as a hierarchical heterogeneous graph consisting of word-level component graphs which introduce more semantic and syntactic information. Then, we dynamically learn a short document graph that facilitates effective label propagation among similar short texts. Thus, compared with existing GNN-based methods, SHINE can better exploit interactions between nodes of the same types and capture similarities between short texts. Extensive experiments on various benchmark short text datasets show that SHINE consistently outperforms state-of-the-art methods, especially with fewer labels.

Gen Shi, Yifan Zhu, Wenjin Liu et al.

Graph neural networks (GNNs) provide powerful insights for brain neuroimaging technology from the view of graphical networks. However, most existing GNN-based models assume that the neuroimaging-produced brain connectome network is a homogeneous graph with single types of nodes and edges. In fact, emerging studies have reported and emphasized the significance of heterogeneity among human brain activities, especially between the two cerebral hemispheres. Thus, homogeneous-structured brain network-based graph methods are insufficient for modelling complicated cerebral activity states. To overcome this problem, in this paper, we present a heterogeneous graph neural network (HebrainGNN) for multimodal brain neuroimaging fusion learning. We first model the brain network as a heterogeneous graph with multitype nodes (i.e., left and right hemispheric nodes) and multitype edges (i.e., intra- and interhemispheric edges). Then, we propose a self-supervised pretraining strategy based on a heterogeneous brain network to address the potential overfitting problem caused by the conflict between a large parameter size and a small medical data sample size. Our results show the superiority of the proposed model over other existing methods in brain-related disease prediction tasks. Ablation experiments show that our heterogeneous graph-based model attaches more importance to hemispheric connections that may be neglected due to their low strength by previous homogeneous graph models. Other experiments also indicate that our proposed model with a pretraining strategy alleviates the problem of limited labelled data and yields a significant improvement in accuracy.

Xiawei Guo, Yuhan Quan, Huan Zhao et al.

Tabular data prediction (TDP) is one of the most popular industrial applications, and various methods have been designed to improve the prediction performance. However, existing works mainly focus on feature interactions and ignore sample relations, e.g., users with the same education level might have a similar ability to repay the debt. In this work, by explicitly and systematically modeling sample relations, we propose a novel framework TabGNN based on recently popular graph neural networks (GNN). Specifically, we firstly construct a multiplex graph to model the multifaceted sample relations, and then design a multiplex graph neural network to learn enhanced representation for each sample. To integrate TabGNN with the tabular solution in our company, we concatenate the learned embeddings and the original ones, which are then fed to prediction models inside the solution. Experiments on eleven TDP datasets from various domains, including classification and regression ones, show that TabGNN can consistently improve the performance compared to the tabular solution AutoFE in 4Paradigm.