High-fidelity agent initialization is crucial for credible Agent-Based Modeling across diverse domains.A robust framework should be Topic-Adaptive, capturing macro-level joint distributions while ensuring micro-level individual rationality.Existing approaches fall into two categories: static data-based retrieval methods that fail to adapt to unseen topics absent from the data, and LLM-based generation methods that lack macro-level distribution awareness, resulting in inconsistencies between micro-level persona attributes and reality.To address these problems, we propose HAG, a Hierarchical Agent Generation framework that formalizes population generation as a two-stage decision process.Firstly, utilizing a World Knowledge Model to infer hierarchical conditional probabilities to construct the Topic-Adaptive Tree, achieving macro-level distribution alignment.Then, grounded real-world data, instantiation and agentic augmentation are carried out to ensure micro-level consistency.Given the lack of specialized evaluation, we establish a multi-domain benchmark and a comprehensive PACE evaluation framework.Extensive experiments show that HAG significantly outperforms representative baselines, reducing population alignment errors by an average of 37.7% and enhancing sociological consistency by 18.8%.

Xiucheng Xu, Bingbing Xu, Tian Xueyun, Zihe Huang, Rongxin Chen, Li Yunfan, Huawei Shen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Tian Xueyun, MingHua Ma, Bingbing Xu, Nuoyan Lyu, Wei Li, Heng Dong, Zheng Chu, Yuanzhuo Wang, Huawei Shen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Deep Research (DR) requires LLM agents to autonomously perform multi-step information seeking, processing, and reasoning to generate comprehensive reports. In contrast to existing studies that mainly focus on unstructured web content, a more challenging DR task should additionally utilize structured knowledge to provide a solid data foundation, facilitate quantitative computation, and lead to in-depth analyses. In this paper, we refer to this novel task as Knowledgeable Deep Research (KDR), which requires DR agents to generate reports with both structured and unstructured knowledge. Furthermore, we propose the Hybrid Knowledge Analysis framework (HKA), a multi-agent architecture that reasons over both kinds of knowledge and integrates the texts, figures, and tables into coherent multimodal reports. The key design is the Structured Knowledge Analyzer, which utilizes both coding and vision-language models to produce figures, tables, and corresponding insights. To support systematic evaluation, we construct KDR-Bench, which covers 9 domains, includes 41 expert-level questions, and incorporates a large number of structured knowledge resources (e.g., 1,252 tables). We further annotate the main conclusions and key points for each question and propose three categories of evaluation metrics including general-purpose, knowledge-centric, and vision-enhanced ones. Experimental results demonstrate that HKA consistently outperforms most existing DR agents on general-purpose and knowledge-centric metrics, and even surpasses the Gemini DR agent on vision-enhanced metrics, highlighting its effectiveness in deep, structure-aware knowledge analysis. Finally, we hope this work can serve as a new foundation for structured knowledge analysis in DR agents and facilitate future multimodal DR studies.

Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. Existing retraining-based approaches are costly and poorly scalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness. In this paper, we propose a novel Multi-Personality Generation (MPG) framework under the decoding-time combination paradigm. It flexibly controls multi-personality without relying on scarce multi-dimensional models or extra training, leveraging implicit density ratios in single-dimensional models as a ''free lunch'' to reformulate the task as sampling from a target strategy aggregating these ratios. To implement MPG efficiently, we design Speculative Chunk-level based Rejection sampling (SCR), which generates responses in chunks and parallelly validates them via estimated thresholds within a sliding window. This significantly reduces computational overhead while maintaining high-quality generation. Experiments on MBTI personality and Role-Playing demonstrate the effectiveness of MPG, showing improvements up to 16%–18%. Code and data are available at https://github.com/Libra117/MPG.

Inference-time alignment methods have gained significant attention for their efficiency and effectiveness in aligning large language models (LLMs) with human preferences.However, existing dominant approaches using rewardguided search (RGS) primarily rely on outcome reward models (ORMs), which suffer from a critical granularity mismatch: ORMs are designed to provide outcome rewards for complete responses, while RGS methods rely on process rewards to guide the policy, leading to inconsistent scoring and suboptimal alignment.To address this challenge, we introduce process reward models (PRMs) into RGS and argue that an ideal PRM should satisfy two objectives: Score Consistency, ensuring coherent evaluation across partial and complete responses, and Preference Consistency, aligning partial sequence assessments with human preferences.Based on these, we propose SP-PRM, a novel dual-consistency framework integrating score consistency-based and preference consistency-based partial evaluation modules without relying on human annotation.Extensive experiments on dialogue, summarization, and reasoning tasks demonstrate that SP-PRM substantially enhances existing RGS methods, achieving a 3.6%-10.3%improvement in GPT-4 evaluation scores across all tasks.Code is publicly available at this link.

Social simulation has long played a crucial role in exploring the mechanisms underlying human behavior and societal structures. Traditional social simulation relies on rule-based or statistical models, which makes it difficult to capture the complexity and variability of the real world. With the emergence and rapid development of large language model (LLM), new frontiers have been opened toward leveraging LLMs as agent to model human behavior and interactions. This cutting-edge direction has gained significant attention and demonstrated promising results, not only advancing research across a wide range of social science disciplines, but also enabling practical applications in role-playing scenarios. However, this field still faces multiple challenges, such as capturing real-world social phenomena, eliminating bias or ethical considerations, and ensuring usability and reliability. This workshop on LLM Agent for Social Simulation (LASS) aims to bring together researchers and practitioners from diverse backgrounds to foster interdisciplinary collaboration, address key challenges, explore new technologies, and chart promising future directions in this rapidly evolving field.

Despite significant progress in diffusion-based image generation, subject-driven generation and instruction-based editing remain challenging. Existing methods typically treat them separately, struggling with limited high-quality data and poor generalization. However, both tasks require capturing complex visual variations while maintaining consistency between inputs and outputs. Inspired by this, we propose MIGE, a unified framework that standardizes task representations using multimodal instructions. It first treats subject-driven generation as creation on a blank canvas and instruction-based editing as modification of an existing image, establishing a shared input-output formulation, then introduces a novel multimodal encoder that maps free-form multimodal instructions into a unified vision-language space, integrating visual and semantic features through a feature fusion mechanism. This unification enables joint training of both tasks, providing two key advantages: (1) Cross-Task Enhancement: by leveraging shared visual and semantic representations, joint training improves instruction adherence and visual consistency in both subject-driven generation and instruction-based editing. (2) Generalization: learning in a unified format facilitates cross-task knowledge transfer, enabling MIGE to generalize to novel compositional tasks, including instruction-based subject-driven editing. Experiments show that MIGE excels in both subject-driven generation and instruction-based editing while setting a SOTA in the new task of instruction-based subject-driven editing. Code and model have been publicly available at https://github.com/Eureka-Maggie/MIGE.

Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51\%}$ on AdvBench and $\textbf{16.36\%}$ on MATH. Code is publicly available at [this link](https://github.com/yuanyige/sea).

As an important graph pre-training method, Graph Contrastive Learning (GCL) continues to play a crucial role in the ongoing surge of research on graph foundation models or LLM as enhancer for graphs. Traditional GCL optimizes InfoNCE by using augmentations to define self-supervised tasks, treating augmented pairs as positive samples and others as negative. However, this leads to semantically similar pairs being classified as negative, causing significant sampling bias and limiting performance. In this paper, we argue that GCL is essentially a Positive-Unlabeled (PU) learning problem, where the definition of self-supervised tasks should be semantically guided, i.e., augmented samples with similar semantics are considered positive, while others, with unknown semantics, are treated as unlabeled. From this perspective, the key lies in how to extract semantic information. To achieve this, we propose IFL-GCL, using InfoNCE as a "free lunch" to extract semantic information. Specifically, We first prove that under InfoNCE, the representation similarity of node pairs aligns with the probability that the corresponding contrastive sample is positive. Then we redefine the maximum likelihood objective based on the corrected samples, leading to a new InfoNCE loss function. Extensive experiments on both the graph pretraining framework and LLM as an enhancer show significantly improvements of IFL-GCL in both IID and OOD scenarios, achieving up to a 9.05% improvement, validating the effectiveness of semantically guided. Code for IFL-GCL is publicly available at: https://github.com/Camel-Prince/IFL-GCL.

Large language models (LLMs) have achieved remarkable progress across various domains, yet their tendency to generate hallucinations remains a critical barrier to their practical reliability.Confidence calibration addresses this challenge by aligning a model's confidence with its actual accuracy, improving self-evaluation and trustworthiness.However, traditional confidence calibration, operating at response level, are inadequate for long-form generation, which involve complex outputs composed of multiple atomic facts, each with varying confidence, correctness, and relevance to the query.To overcome this limitation, we propose a fact-level confidence calibration framework that evaluates and adjusts confidence at the granularity of individual facts, incorporating both relevance and correctness.This framework identifies finer-grained calibration discrepancies, reduces overconfidence, and reveals confidence variance.Based on this framework, we introduce CARE (Confidence-Aware Fact Correction), a method that leverages high-confidence facts to iteratively refine and correct low-confidence ones.Experimental results demonstrate that our CARE effectively improves the quality of generated content.Our code is available at this link.

The generalization of neural networks is a central challenge in machine learning, especially concerning the performance under distributions that differ from training ones. Current methods, mainly based on the data-driven paradigm such as data augmentation, adversarial training, and noise injection, may encounter limited generalization due to model non-smoothness. In this paper, we propose to investigate generalization from a Partial Differential Equation (PDE) perspective, aiming to enhance it directly through the underlying function of neural networks, rather than focusing on adjusting input data. Specifically, we first establish the connection between neural network generalization and the smoothness of the solution to a specific PDE, namely transport equation. Building upon this, we propose a general framework that introduces adaptive distributional diffusion into transport equation to enhance the smoothness of its solution, thereby improving generalization. In the context of neural networks, we put this theoretical framework into practice as PDE+ (PDE with Adaptive Distributional Diffusion) which diffuses each sample into a distribution covering semantically similar inputs. This enables better coverage of potentially unobserved distributions in training, thus improving generalization beyond merely data-driven methods. The effectiveness of PDE+ is validated through extensive experimental settings, demonstrating its superior performance compared to state-of-the-art methods. Our code is available at https://github.com/yuanyige/pde-add.

Test Time Adaptation (TTA) aims to improve model generalizability when test data diverges from training distribution, with the distinct advantage of not requiring access to training data and processes, especially valuable in the context of pre-trained models. However, current TTA methods fail to address the fundamental issue: covariate shift, i. e., the decreased generalizability can be attributed to the model's reliance on the marginal distribution of the training data, which may impair model calibration and introduce confirmation bias. To address this, we propose a novel energy-based perspective, enhancing the model's perception of target data distributions without requiring access to training data or processes. Building on this perspective, we introduce Test-time Energy Adaptation (TEA), which transforms the trained classifier into an energy-based model and aligns the model's distribution with the test data's, enhancing its ability to perceive test distributions and thus improving overall generalizability. Extensive experiments across multiple tasks, benchmarks and architectures demonstrate TEA's superior generalization performance against state-of-the-art methods. Further in-depth analyses reveal that TEA can equip the model with a comprehensive perception of test distribution, ultimately paving the way toward improved generalization and calibration11Code is available at https://github.com/yuanyige/tea

Community search is a personalized community discovery problem aimed at finding densely-connected subgraphs containing the query vertex. In particular, the search for com-munities with high-importance vertices has recently received a great deal of attention. However, existing works mainly focus on conventional homogeneous networks where vertices are of the same type, but are not applicable to heterogeneous information networks (HINs) composed of multi-typed vertices and different semantic relations, such as bibliographic networks. In this paper, we study the problem of high-importance community search in HINs. A novel community model is introduced, named heterogeneous significant community (HSC), to unravel the closely connected vertices of the same type with high attribute values through multiple semantic relationships. An HSC not only maximizes the exploration of indirect relationships across entities of the anchor-type but incorporates their significance. To search the HSCs, we first develop online algorithms by exploiting both segmented-based meta-path expansion and significance incrernent. Specially, a solution space reuse strategy based on structural nesting is designed to boost the efficiency. In addition, we further devise a two-level index to support searching HSCs in optimal time, based on which a space-efficient compact index is proposed. Extensive experiments on real-world large-scale HINs demonstrate that our solutions are effective and efficient for searching HSCs, and the index-based algorithms are 2–4 orders of magnitude faster than online algorithms.

Graph contrastive learning (GCL), standing as the dominant paradigm in the realm of graph pre-training, has yielded considerable progress. Nonetheless, its capacity for out-of-distribution (OOD) generalization has been relatively underexplored. In this work, we point out that the traditional optimization of InfoNCE in GCL restricts the cross-domain pairs only to be negative samples, which inevitably enlarges the distribution gap between different domains. This violates the requirement of domain invariance under OOD scenario and consequently impairs the model's OOD generalization performance. To address this issue, we propose a novel strategy ''Negative as Positive'', where the most semantically similar cross-domain negative pairs are treated as positive during GCL. Our experimental results, spanning a wide array of datasets, confirm that this method substantially improves the OOD generalization performance of GCL.

The data-hungry problem, characterized by insufficiency and low-quality of data, poses obstacles for deep learning models. Transfer learning has been a feasible way to transfer knowledge from high-quality external data of source domains to limited data of target domains, which follows a domain-level knowledge transfer to learn a shared posterior distribution. However, they are usually built on strong assumptions, e.g., the domain invariant posterior distribution, which is usually unsatisfied and may introduce noises, resulting in poor generalization ability on target domains. Inspired by Graph Neural Networks (GNNs) that aggregate information from neighboring nodes, we redefine the paradigm as learning a knowledge-enhanced posterior distribution for target domains, namely Knowledge Bridge Learning (KBL). KBL first learns the scope of knowledge transfer by constructing a Bridged-Graph that connects knowledgeable samples to each target sample and then performs sample-wise knowledge transfer via GNNs.KBL is free from strong assumptions and is robust to noises in the source data. Guided by KBL, we propose the Bridged-GNN including an Adaptive Knowledge Retrieval module to build Bridged-Graph and a Graph Knowledge Transfer module. Comprehensive experiments on both un-relational and relational data-hungry scenarios demonstrate the significant improvements of Bridged-GNN compared with SOTA methods

The importance of social bot detection has been increasingly recognized due to its profound impact on information dissemination. Existing methodologies can be categorized into feature engineering and deep learning-based methods, which mainly focus on static features, e.g., post characteristics and user profiles.However, existing methods often overlook the burst phenomena when distinguishing social bots and genuine users, i.e, the sudden and intense activity or behavior of bots after prolonged inter. Through comprehensive analysis, we find that both burst behavior and static features play pivotal roles in social bot detection. To capture such properties, the dual-channel GNN (DCGNN) is proposed which consists of a burst-aware channel with an adaptive-pass filter and a static-aware channel with a low-pass filter to model user characteristics effectively. Experimental results demonstrate the superiority of this method over competitive baselines.

Traffic forecasting is crucial for public safety and resource optimization, yet is very challenging due to the temporal changes and the dynamic spatial correlations of the traffic data. To capture these intricate dependencies, spatio-temporal networks, such as recurrent neural networks with graph convolution networks, graph convolution networks with temporal convolution networks, and temporal attention networks with full graph attention networks, are applied. However, previous spatio-temporal networks are based on end-to-end training and thus fail to handle the distribution shift in the non-stationary traffic time series. On the other hand, the efficient and effective algorithm for modeling spatial correlations is still lacking in prior networks.In this paper, rather than proposing yet another end-to-end model, we aim to provide a novel disentangle-fusion framework STWave to mitigate the distribution shift issue. The framework first decouples the complex traffic data into stable trends and fluctuating events, followed by a dual-channel spatio-temporal network to model trends and events, respectively. Finally, reasonable future traffic can be predicted through the fusion of trends and events. Besides, we incorporate a novel query sampling strategy and graph wavelet-based graph positional encoding into the full graph attention network to efficiently and effectively model dynamic spatial correlations. Extensive experiments on six traffic datasets show the superiority of our approach, i.e., the higher forecasting accuracy with lower computational cost.

Unsupervised representation learning for dynamic graphs has attracted a lot of research attention in recent years. Compared with static graph, the dynamic graph is a comprehensive embodiment of both the intrinsic stable characteristics of nodes and the time-related dynamic preference. However, existing methods generally mix these two types of information into a single representation space, which may lead to poor explanation, less robustness, and a limited ability when applied to different downstream tasks. To solve the above problems, in this paper, we propose a novel disenTangled representation learning framework for discrete-time Dynamic graphs, namely DyTed. We specially design a temporal-clips contrastive learning task together with a structure contrastive learning to effectively identify the time-invariant and time-varying representations respectively. To further enhance the disentanglement of these two types of representation, we propose a disentanglement-aware discriminator under an adversarial learning framework from the perspective of information theory. Extensive experiments on Tencent and five commonly used public datasets demonstrate that DyTed, as a general framework that can be applied to existing methods, achieves state-of-the-art performance on various downstream tasks, as well as be more robust against noise.

"Transaction fraud" complaint verification, i.e., verifying whether a transaction corresponding to a complaint is fraudulent, is particularly critical to prevent economic loss. Compared with traditional fraud pre-transaction detection, complaint verification puts forward higher requirements: 1)an individual tends to exhibit different identities in different complaints, e.g., complainant or respondent, requiring the model to capture identity-related representations corresponding to the complaint; 2)the fraud ways evolve frequently to confront detection, requiring the model to perform stably under different fraud ways. Previous methods mainly focused on fraud pre-transaction detection, utilizing the historical information of users or conduct message passing based GNNs on relationship networks. However, they rarely consider capturing various identity-related representations and ignore the evolution of fraud ways, leading to failure in complaint verification. To address the above challenges, we propose the mutual information based dual level graph neural network, namely MIDLG, which defines a complaint as a super-node consisting of involved individuals, and characterizes the individual over node-level and super-node-level. Furthermore, the mutual information minimization objective is proposed based on "complaint verification-causal graph" to decouple the model prediction from relying on specific fraud ways, and thus achieve stability. MIDLG achieves SOTA results through extensive experiments in complaint verification on WeChat Finance, one online payment service serving more than 600 million users in China.

Graphs consisting of vocal nodes ("the vocal minority") and silent nodes ("the silent majority"), namely VS-Graph, are ubiquitous in the real world. The vocal nodes tend to have abundant features and labels. In contrast, silent nodes only have incomplete features and rare labels, e.g., the description and political tendency of politicians (vocal) are abundant while not for ordinary civilians (silent) on the twitter’s social network. Predicting the silent majority remains a crucial yet challenging problem. However, most existing Graph Neural Networks (GNNs) assume that all nodes belong to the same domain, without considering the missing features and distribution-shift between domains, leading to poor ability to deal with VS-Graph. To combat the above challenges, we propose Knowledge Transferable Graph Neural Network (KTGNN), which models distribution-shifts during message passing and learns representation by transferring knowledge from vocal nodes to silent nodes. Specifically, we design the domain-adapted "feature completion and message passing mechanism" for node representation learning while preserving domain difference. And a knowledge transferable classifier based on KL-divergence is followed. Comprehensive experiments on real-world scenarios (i.e., company financial risk assessment and political elections) demonstrate the superior performance of our method. Our source code has been open-sourced1.

Contrastive learning gains remarkable success in graph representation learning through imposing two properties to node representations, i.e., aligning representations of similar nodes in different augmented graphs and uniformly distributing node representations on the hypersphere. Existing graph contrastive learning methods capture these two key properties by extensive negative examples, resulting in unbearable computational costs. To overcome this issue, we propose a novel graph contrastive learning method, i.e., GraphALU, to capture these two properties without negative examples. For alignment, we leverage both self-alignment, i.e., pulling the representations of the same node in different augmented graphs together, and neighbor-alignment, pulling the representations of neighboring nodes in different augmented graphs together to capture the structure proximity among nodes. For uniformity, we innovatively design a star-shaped distance comparison, i.e., convert the distance comparison of any pair of nodes to the distance comparison of any nodes with a virtual center node, reducing the quadratic computational cost to a linear one with respect to the number of nodes. Extensive experiments show both the effectiveness and efficiency of GraphALU on various established benchmark datasets and downstream tasks.

For the problem that GNSS can't be used for positioning indoors and with the development of computing technology and the deployment widespread of Wi-Fi, Wi-Fi based indoor crowdsourced positioning methods have gradually become a hot spot for research and application. However, compared with professional fingerprint acquisition, crowdsourced data lacks accurate indoor labeled locations, which creates a serious positioning error if used to build the fingerprint database directly. Most of the current methods add too many constraints. To solve above issues, this paper proposes a method for the construction of crowdsourced location fingerprint map, which solves the error problem of data point dislocation caused by manifold dimensionality reduction while using only a small amount of labeled data which is some contain real physical coordinates data points. Using these data points as reference points and using some environmental elements constraining each data point to complete the build of the entire indoor environment. Practical experiment shows that the error of the proposed method can be reduced to 6m.

Company financial risk is ubiquitous and early risk assessment for listed companies can avoid considerable losses. Traditional methods mainly focus on the financial statements of companies and lack the complex relationships among them. However, the financial statements are often biased and lagged, making it difficult to identify risks accurately and timely. To address the challenges, we redefine the problem as company financial risk assessment on tribe-style graph by taking each listed company and its shareholders as a tribe and leveraging financial news to build inter-tribe connections. Such tribe-style graphs present different patterns to distinguish risky companies from normal ones. However, most nodes in the tribe-style graph lack attributes, making it difficult to directly adopt existing graph learning methods (e.g., Graph Neural Networks(GNNs)). In this paper, we propose a novel Hierarchical Graph Neural Network (TH-GNN) for Tribe-style graphs via two levels, with the first level to encode the structure pattern of the tribes with contrastive learning, and the second level to diffuse information based on the inter-tribe relations, achieving effective and efficient risk assessment. Extensive experiments on the real-world company dataset show that our method achieves significant improvements on financial risk assessment over previous competing methods. Also, the extensive ablation studies and visualization comprehensively show the effectiveness of our method.

Consumer loans, i.e., loans to finance consumers to buy certain types of expenditures, is increasingly popular in e-commerce platform. Different from traditional loans with mortgage, online consumer loans only take personal credit as collateral for loans. Consequently, loan fraud detection is particularly critical for lenders to avoid economic loss. Previous methods mainly leverage applicant's attributes and historical behavior for loan fraud detection. Although these methods gain success at detecting potential charge-offs, yet they perform worse when multiple persons with various roles (e.g., sellers, intermediaries) collude to apply fraudulent loan. To combat this challenge, we consider the problem of loan fraud detection via exploiting roles of users and multi-type social relationships among users. We propose a novel Graph neural network with a Role-constrained Conditional random field, namely GRC, to learn the representation of applicants and detect loan fraud based on the learned representation. The proposed model characterizes the multiple types of relationships via self-attention mechanism and employs conditional random field to constrain users with the same role to have similar representation. We validate the proposed model through experiments in large-scale auto-loan scenario. Extensive experiments demonstrate that our model achieves state-of-the-art results in loan fraud detection on Alipay, one online credit payment service serving more than 450 million users in China.

Graph neural networks (GNNs) achieve remarkable success in graph-based semi-supervised node classification, leveraging the information from neighboring nodes to improve the representation learning of target node. The success of GNNs at node classification depends on the assumption that connected nodes tend to have the same label. However, such an assumption does not always work, limiting the performance of GNNs at node classification. In this paper, we propose label-consistency based graph neural network (LC-GNN), leveraging node pairs unconnected but with the same labels to enlarge the receptive field of nodes in GNNs. Experiments on benchmark datasets demonstrate the proposed LC-GNN outperforms traditional GNNs in graph-based semi-supervised node classification. We further show the superiority of LC-GNN in sparse scenarios with only a handful of labeled nodes.

Graph convolutional networks gain remarkable success in semi-supervised learning on graph-structured data. The key to graph-based semisupervised learning is capturing the smoothness of labels or features over nodes exerted by graph structure. Previous methods, spectral methods and spatial methods, devote to defining graph convolution as a weighted average over neighboring nodes, and then learn graph convolution kernels to leverage the smoothness to improve the performance of graph-based semi-supervised learning. One open challenge is how to determine appropriate neighborhood that reflects relevant information of smoothness manifested in graph structure. In this paper, we propose GraphHeat, leveraging heat kernel to enhance low-frequency filters and enforce smoothness in the signal variation on the graph. GraphHeat leverages the local structure of target node under heat diffusion to determine its neighboring nodes flexibly, without the constraint of order suffered by previous methods. GraphHeat achieves state-of-the-art results in the task of graph-based semi-supervised classification across three benchmark datasets: Cora, Citeseer and Pubmed.