论文检索

Lei Wang 0197, Shanshan Huang 0004, Chunyuan Zheng 0001, Jun Liao 0001, Xiaofei Zhu, Haoxuan Li 0001, Li Liu 0001

Imbalanced Time-Series Classification is a critical, yet challenging task across a spectrum of real-world applications. Previous oversampling and generative approaches primarily target the minority class and often rely on static decision boundaries or similarity-based heuristics. However, these methods overlook the underlying causal factors that govern the distinction between majority and minority classes, particularly in scenarios with ambiguous class boundaries. As a result, the generated samples may fail to enhance class separability, thereby limiting improvements in classification performance. To this end, we propose a CounterFactual Augmentation Minority Generation (CFAMG) method based on generative models that aims to discover the causal factors that determine different classes from a causality perspective. Specifically, our method first utilizes a disentangled classifier to distinguish between causal and non-causal factors. Next, we perform counterfactual intervention by replacing the causal factors of majority class samples with those from minority class samples, creating an intervened latent representation that reflects minority characteristics while preserving essential structures. Finally, the trained minority class decoder generates counterfactual minority samples that resemble real minority instances yet remain distinguishable from the original majority class. Extensive experiments demonstrate that our method outperforms state-of-the-art methods in both univariate and multivariate imbalanced time-series classification tasks. The code is published at https://github.com/WangLei-CQU/CFAMG.

Chenxu Wang 0001, Jinfeng Chen, Junzhou Zhao, Pinghui Wang

Graph Few-Shot Node Classification (GFSNC) has emerged as a promising approach to address the challenge of learning with limited labeled data in graph-structured networks. Despite the success of Graph Neural Networks (GNNs) in node classification tasks, their performance heavily depends on the availability of abundant labeled data, which is often impractical in real-world scenarios. To tackle this issue, GFSNC adopts the episodic meta-learning paradigm, where models are trained on a series of meta-tasks. However, existing methods face two critical limitations: (i) they focus on local distributions within individual meta-tasks, neglecting the global data distribution, and (ii) they optimize models to minimize intra-class distances without adequately addressing inter-class separability, leading to suboptimal performance. This paper presents TaskNS, a novel GFSNC framework that introduces task-negative samples into meta-training tasks to address these limitations. By incorporating samples from classes outside the current meta-task, our framework enables the model to gradually learn the global distribution of the graph data. Additionally, we design a novel loss function that enhances the model's ability to distinguish between different classes of query samples. This loss function not only ensures high intra-class compactness but also maximizes the inter-class separation by leveraging task-negative samples. To further enhance the quality of task-negative samples, we propose an h-hop-neighbors-based sampling method that leverages the topological structure of a graph. It selects task-negative samples that are structurally close to query samples, ensuring that they are informative and challenging for the model to classify. Extensive experiments on four benchmark datasets demonstrate the effectiveness of TaskNS, achieving average improvements of 4.6% in accuracy (ACC) and 4.9% in F1-score over state-of-the-art methods.

Nilesh Verma, Albert Bifet, Bernhard Pfahringer, Maroua Bahri

Hyperparameter optimization is crucial for maximizing machine learning model performance, yet most existing algorithms are designed for batch or offline scenarios and assume static data distributions. Such assumptions fall short in data stream settings, where models must adapt to evolving inputs in real time. To address these limitations, we propose the Bayesian Stream Tuner (BST), a novel framework for online hyperparameter optimization in non-stationary data streams. BST maintains a dynamic set of candidate hyperparameter configurations and periodically refines them using an incremental Bayesian model, which estimates configuration performance based on recent data statistics and hyperparameter values. This systematic exploration and refinement strategy allows BST to detect and respond to concept drift by resetting its adaptation mechanisms whenever necessary, ensuring strong performance under changing distributions. Our theoretical analysis establishes sublinear regret bounds for BST in dynamic environments, and extensive experiments on classification and regression tasks demonstrate that BST consistently outperforms state-of-the-art online hyperparameter optimization methods in both predictive accuracy and adaptability, making it a powerful solution for real-time hyperparameter tuning in evolving data streams.

Ram Ganesh V, Ayush Singh, Aditi Rai, Harsh Pal, Deepanshu Bagotia, Akshay Sethi, Aakarsh Malhotra, Sayan Ranu

Graph Neural Networks (Gnns) have achieved remarkable success in various downstream tasks, such as node classification and link prediction. Yet, efficiently deploying Gnns remains a challenge due to their computational complexity. Graph knowledge distillation aims to address this by transferring task-specific structural knowledge from teacher Gnns to lightweight student Gnns or Multi-Layer Perceptrons (MLPs). Despite its promise, existing distillation approaches suffer from several limitations: (i) they require extensive task-specific supervision(ii) they must be retrained separately for each downstream task, and (iii) they often struggle in heterophilous settings. To overcome these challenges, we propose TAG2M, a Task-Agnostic Gnn-to-MLP distillation framework designed for efficient and accurate few-shot inference. TAG2M introduces several novel strategies, including a self-supervised contrastive loss that captures topological information solely from node attributes. Additionally, it leverages Lipschitz embeddings to encode positional information with provable distortion bounds, ensuring robust representation learning. To further enhance adaptability for few-shot inference, TAG2M incorporates a learnable prompt head, which facilitates rapid task adaptation even in label-scarce settings. Unlike prior methods, TAG2M generalizes well across both homophilous and heterophilous datasets while delivering a significant computational advantage, achieving up to a 20X -200X speed-up. Extensive evaluations on 11 public datasets demonstrate its superior accuracy across diverse tasks, including node classification, link prediction, and node regression, outperforming state-of-the-art approaches.

Tian Tian 0008, Chunyan Miao, Hangwei Qian

Contrastive learning has emerged as a competent approach for unsupervised representation learning. However, the design of an optimal augmentation strategy, although crucial for contrastive learning, is less explored for time series classification tasks. Existing predefined time-domain augmentation methods are primarily adopted from vision and are not specific to time series data. Consequently, this cross-modality incompatibility may distort the semantically relevant information of time series by introducing mismatched patterns into the data. To address this limitation, we present a novel perspective from the frequency domain and identify three advantages for downstream classification: 1) the frequency component naturally encodes global features, 2) the orthogonal nature of the Fourier basis allows easier isolation and independent modifications of critical and unimportant information, and 3) a compact set of frequency components can preserve semantic integrity. To fully utilize the three properties, we propose the lightweight yet effective Frequency-Refined Augmentation (FreRA) tailored for time series contrastive learning on classification tasks, which can be seamlessly integrated with contrastive learning frameworks in a plug-and-play manner. Specifically, FreRA automatically separates critical and unimportant frequency components. Accordingly, we propose semantic-aware Identity Modification and semantic-agnostic Self-adaptive Modification to protect semantically relevant information in the critical frequency components and infuse variance into the unimportant ones respectively. Theoretically, we prove that FreRA generates semantic-preserving views. Empirically, we conduct extensive experiments on two benchmark datasets, including UCR and UEA archives, as well as five large-scale datasets on diverse applications. FreRA consistently outperforms ten leading baselines on time series classification, anomaly detection, and transfer learning tasks, demonstrating superior capabilities in contrastive representation learning and generalization in transfer learning scenarios across diverse datasets. The code is available at https://github.com/Tian0426/FreRA.

Yu Song 0007, Zhigang Hua, Harry Shomer, Yan Xie, Jingzhe Liu, Bo Long, Hui Liu 0031

Link Prediction (LP) is a critical task in graph machine learning. While Graph Neural Networks (GNNs) have significantly advanced LP performance recently, existing methods face key challenges including limited supervision from sparse connectivity, sensitivity to initialization, and poor generalization under distribution shifts. We explore pretraining as a solution to address these challenges. Unlike node classification, LP is inherently a pairwise task, which requires the integration of both node- and edge-level information. In this work, we present the first systematic study on the transferability of these distinct modules and propose a late fusion strategy to effectively combine their outputs for improved performance. To handle the diversity of pretraining data and avoid negative transfer, we introduce a Mixture-of-Experts (MoE) framework that captures distinct patterns in separate experts, facilitating seamless application of the pretrained model on diverse downstream datasets. For fast adaptation, we develop a parameter-efficient tuning strategy that allows the pretrained model to adapt to unseen datasets with minimal computational overhead. Experiments on 16 datasets across two domains demonstrate the effectiveness of our approach, achieving state-of-the-art performance on low-resource link prediction while obtaining competitive results compared to end-to-end trained methods, with over 10,000x lower computational overhead.

Zeang Sheng, Weiyang Guo, Yingxia Shao, Wentao Zhang 0001, Bin Cui 0001

Graph Neural Networks (GNNs) have achieved significant success in various graph learning tasks. However, their success relies heavily on the availability of adequate high-quality labels, which requires high annotation costs. Recently, Large Language Models (LLMs) became known to the community for their strong zero-shot performance on diverse textual tasks. To utilize LLM's zero-shot abilities, existing work substitutes the oracle in the graph active learning setting with an LLM and achieves low-cost annotation compared to consulting human experts. However, annotations from LLMs inevitably contain noise, and we find through an empirical analysis that LLMs possess complex and dataset-specific annotation noise distributions. This intricacy of LLMs' annotation noise is neglected by existing work, leading to sub-optimal annotation data quality under the same budget. In this paper, we propose a Dataset- and LLM-Aware graph active learning framework called DMA that explicitly models LLM's annotation noise distributions on the given dataset. Concretely, DMA queries the oracle LLM with manually constructed prompts for each class's pseudo samples that are then used to measure the class-wise annotation noise rates. We then equip DMA with a new noise-aware graph active learning algorithm to utilize this fine-grained annotation noise distribution. Extensive performance optimizations are applied to the implementation to boost DMA's scalability and runtime efficiency on large-scale datasets. Evaluations on five text-attributed graph datasets show that DMA consistently outperforms all the baselines in terms of annotation data quality, which illustrates the importance of careful handling of LLM's intricate annotation noise.

Masoud Reyhani Hamedani, Jeong-Seok Oh, Seong-Un Cho, Sang-Wook Kim

In the literature, various graph embedding methods have been proposed. Although they have pioneered notable techniques in the field, we point out their four drawbacks as follows: (1) inability to consider global graph structure(2) undermining learning quality(3) impairing in/out-degree distributions in directed graphs, and (4) limited applicability. Inspired by these drawbacks, we first propose LINOW, a recursive LI nk-based similarity measure for graphs by utilizing NO des' Weights, which is applicable to both directed and undirected graphs. Then, we provide a matrix form that dramatically accelerates LINOW's computation without approximation. Furthermore, to enhance its scalability, we provide two variants, LINOW-sn and LINOW-bn, to compute similarity scores w.r.t. a single node and a batch of nodes, respectively. Finally, we propose SIGEM, a simple yet effective self-supervised and contrastive-free SI milarity based Graph EM bedding method that employs LINOW-bn to compute similarity scores of nodes in the graph, thereby ranking them. Then, it tries to preserve the original ranks of nodes in the graph within their corresponding vectors in the embedding space, by employing a single-layer neural network. The results of our extensive experiments with eight real-world datasets and thirteen state-of-the-art and conventional embedding methods demonstrate that (1) LINOW-sn and LINOW-bn successfully improve the scalability of naive LINOW(2) LINOW is beneficial to similarity based graph embedding, and (3) SIGEM consistently achieves the highest accuracy in both graph reconstruction and node classification tasks compared to other methods, while it significantly outperforms them in most cases of the link prediction task.

Dazhuo Qiu, Jinwen Chen, Arijit Khan 0001, Yan Zhao 0008, Francesco Bonchi

Counterfactual learning is emerging as an important paradigm, rooted in causality, which promises to alleviate common issues of graph neural networks (GNNs), such as fairness and interpretability. However, as in many real-world application domains where conducting randomized controlled trials is impractical, one has to rely on available observational (factual) data to detect counterfactuals. In this paper, we introduce and tackle the problem of searching for counterfactual evidences for the GNN-based node classification task. A counterfactual evidence is a pair of nodes such that, regardless they exhibit great similarity both in the features and in their neighborhood subgraph structures, they are classified differently by the GNN. We develop effective and efficient search algorithms and a novel indexing solution that leverages both node features and structural information to identify counterfactual evidences, and generalizes beyond any specific GNN. Through various downstream applications, we demonstrate the potential of counterfactual evidences to enhance fairness and accuracy of GNNs.

Nabil Anan Orka, Ehtashamul Haque, Md. Abdul Awal, Mohammad Ali Moni

Despite the advancements in quantum convolution or quanvolution, challenges persist in making quanvolution scalable, efficient, and applicable to multi-dimensional data. Existing quanvolutional networks heavily rely on classical layers, with minimal quantum involvement due to inherent limitations in current quanvolution algorithms. Moreover, the application of quanvolution in the domain of 1D data remains largely unexplored. To address these limitations, we propose a new quanvolution algorithm-Quanv1D-capable of processing arbitrary-channel 1D data, handling variable kernel sizes, and generating a customizable number of feature maps, along with a classification network-fully quanvolutional network (FQN)-built solely using Quanv1D layers. Quanv1D is inspired by the classical Conv1D and stands out from the quanvolution literature by being fully trainable, modular, and freely scalable with a self-regularizing feature. To evaluate FQN, we tested it on 20 UEA and UCR time series datasets, both univariate and multivariate, and benchmarked its performance against state-of-the-art convolutional models (both quantum and classical). We found FQN to outperform all compared models in terms of average accuracy while using significantly fewer parameters. Additionally, to assess the viability of FQN on real hardware, we conducted a shot-based analysis across all the datasets to simulate statistical quantum noise and found our model robust and equally efficient.

Kohei Obata, Yasuko Matsubara, Yasushi Sakurai

Unsupervised anomaly detection in time series has been a pivotal research area for decades. Current mainstream approaches focus on learning normality, on the assumption that all or most of the samples in the training set are normal. However, anomalies in the training set (i.e., anomaly contamination) can be misleading. Recent studies employ data augmentation to generate pseudo-anomalies and learn the boundary separating the training samples from the augmented samples. Although this approach mitigates anomaly contamination if augmented samples mimic unseen real anomalies, it suffers from several limitations. (1) Covering a wide range of time series anomalies is challenging. (2) It disregards augmented samples that resemble normal samples (i.e., false anomalies). (3) It places too much trust in the labels of training and augmented samples. In response, we propose RedLamp, which employs diverse data augmentations to generate multiclass pseudo-anomalies and learns the multiclass boundary. Such multiclass pseudo-anomalies cover a wide variety of time series anomalies. We conduct multiclass classification using soft labels, which prevents the model from being overconfident and ensures its robustness against contaminated/false anomalies. The learned latent space is inherently explainable as it is trained to separate pseudo-anomalies into multiclasses. Extensive experiments demonstrate the effectiveness of RedLamp in anomaly detection and its robustness against anomaly contamination.

Yang Liu 0348, Deyu Bo, Wenxuan Cao, Yuan Fang 0001, Yawen Li 0001, Chuan Shi 0001

Graph self-supervised learning seeks to learn effective graph representations without relying on labeled data. Among various approaches, graph autoencoders (GAEs) have gained significant attention for their efficiency and scalability. Typically, GAEs take incomplete graphs as input and predict missing elements, such as masked node features or edges. Although effective, our experimental investigation reveals that traditional feature or edge masking paradigms primarily capture low-frequency signals in the graph and fail to learn expressive structural information. To address these issues, we propose Graph Positional Autoencoders (GraphPAE), which employ a dual-path architecture to reconstruct both node features and positions. Specifically, the feature path uses positional encoding to enhance the message-passing processing, improving the GAEs' ability to predict the corrupted information. The position path, on the other hand, leverages node representations to refine positions and approximate eigenvectors, thereby enabling the encoder to learn diverse frequency information. We conduct extensive experiments to verify the effectiveness of GraphPAE, including heterophilic node classification, graph property prediction, and transfer learning. The results demonstrate that GraphPAE achieves state-of-the-art performance and consistently outperforms the baselines by a large margin.

Yijun Li 0002, Jiawei Huang 0001, Jingling Liu, Zhaoyi Li, Wanchun Jiang, Jianxin Wang 0001

Distributed deep learning has been widely deployed in data centers to provide various services such as image classification and speech recognition. To reduce the training time, Top-k compression has become one of the most popular solutions used to shrink the data volume of gradients. Nevertheless, we observe that existing Top-k compression solutions are inefficient when used for large-scale distributed training due to gradient build-up, missing of Top-k gradients, and high compression overhead at the end hosts. To address these problems, we propose SwitchTop-k, which improves the accuracy of selecting Top-k values while ensuring a high compression rate and zero compression overhead. Specifically, SwitchTop-k offloads the Top-k compression from the end hosts to the programmable switches, thus alleviating the gradient build-up and compression overhead. Meanwhile, we propose a sketch-based solution to achieve high accuracy in selecting global Top-k gradients. We also co-design switch logic and end host logic to improve communication efficiency of uncompressed traffic. Finally, we implement SwitchTop-k on Intel Tofino switches and integrate it with Pytorch. The test results show that SwitchTop-k reduces iteration time by up to 91% compared with existing compression algorithms.

Xiaopeng Li 0006, Shasha Li 0001, Shangwen Wang, Shezheng Song, Bin Ji 0002, Huijun Liu 0003, Jun Ma 0015, Jie Yu 0008

Warning: This paper contains examples of toxic text. Knowledge editing has emerged as an efficient technique for updating the knowledge of large language models (LLMs), attracting increasing attention in recent years. However, there is a lack of effective measures to prevent the malicious misuse of this technique, which could lead to harmful edits in LLMs. These malicious modifications could cause LLMs to generate toxic content, misleading users into inappropriate actions. In front of this risk, we introduce a new task, Knowledge Editing Type Identification (KETI), aimed at identifying different types of edits in LLMs, thereby providing timely alerts to users when encountering illicit edits. As part of this task, we propose KETIBench, which includes five types of harmful edits covering the most popular toxic types, as well as one benign factual edit. We develop five classical classification models and three BERT-based models as baseline identifiers for both open-source and closed-source LLMs. Our experimental results, across 92 trials involving four models and three knowledge editing methods, demonstrate that all eight baseline identifiers achieve decent identification performance, highlighting the feasibility of identifying malicious edits in LLMs. Additional analyses reveal that the performance of the identifiers is independent of the reliability of the knowledge editing methods and exhibits cross-domain generalization, enabling the identification of edits from unknown sources. All data and code are available in https://github.com/xpq-tech/KETI.

Tong Li 0017, Jiachuan Wang, Yongqi Zhang, Shuangyin Li, Lei Chen 0002

Citation classification, which identifies the intention behind academic citations, is pivotal for scholarly analysis. Previous works suggest fine-tuning pretrained language models (PLMs) on citation classification datasets, reaping the reward of the linguistic knowledge they gained during pretraining. However, directly fine-tuning for citation classification is challenging due to labeled data scarcity, contextual noise, and spurious keyphrase correlations. In this paper, we present a novel framework, Citss, that adapts the PLMs to overcome these challenges. Citss introduces self-supervised contrastive learning to alleviate data scarcity, and is equipped with two specialized strategies to obtain the contrastive pairs: sentence-level cropping, which enhances focus on target citations within long contexts, and keyphrase perturbation, which mitigates reliance on specific keyphrases. Compared with previous works that are only designed for encoder-based PLMs, Citss is carefully developed to be compatible with both encoder-based PLMs and decoder-based LLMs, to embrace the benefits of enlarged pretraining. Experiments with three benchmark datasets with both encoder-based PLMs and decoder-based LLMs demonstrate our superiority compared to the previous state of the art. Our code is available at: github.com/LITONG99/Citss

Shuangyin Li, Jihua Yang, Yixuan Wang, Shimin Di, Lei Chen 0002

Single-cell RNA sequencing (scRNA-seq) technology has emerged as a powerful tool for studying complex biological systems and diseases with high resolution. However, generating sufficient high quality scRNA-seq samples for accurate downstream analyses and ensuring reproducibility remains a challenge. Previous methods often produce low-quality samples or a limited number of useful specific cell subpopulations. To address this issue, we propose a novel Efficient Single-Cell RNA-seq Flow Diffusion (named ESCFD) method based on the Diffusion Model, capable of synthesizing large-scale, high-quality scRNA-seq samples and generating either ''holistic'' or specific cellular subpopulations within a single framework. A pre-guidance mechanism is designed for synthesizing specific cellular subpopulations of high quality, and a post-guidance mechanism was designed to accelerate the sampling of scRNA-seq samples. The ESCFD can efficiently synthesize large-scale and high-quality scRNA-seq samples for various downstream tasks, surpassing previous work with faster sampling speed. Experimental results demonstrate that the ESCFD model achieves state-of-the-art performance in cell classification and cell distribution similarity when evaluated on three scRNA-seq benchmarks, and verified the effectiveness of accelerated sampling. Visualization experiments also reveal the ESCFD's ability to synthesize specific cellular subpopulations. This innovative approach has the potential to significantly advance our understanding of complex biological systems and accelerate the development of novel therapeutic interventions and personalized medicine strategies.

Ruikun Li 0001, Ye Xiao, Xiaoxiao Ma 0002, Andrey Vasnev, Junbin Gao

A significant challenge for Graph Neural Networks (GNNs) lies in accurately performing node classification tasks on heterophilous graphs. Traditional GNN models, which largely rely on the assumption of homophily, often struggle under heterophilous conditions due to the inherent low-pass filtering nature during message-passing among neighboring nodes. Motivated by these limitations, our research explores the dendritic structures found in real neurons, which dynamically amplify and inhibit signals. We demonstrate how the biophysical properties of dendrites enable GNNs to adaptively modulate the Dirichlet energy of graph representations based on the specific context, providing a theoretical justification for integrating neural dendrites into GNNs to improve learning on heterophilous graphs. In this work, we introduce the GDendrite framework, which dynamically adjusts GNN's output response to the various contexts. Our empirical studies, conducted across 13 homophilous and heterophilous benchmark datasets, show that our framework significantly enhances the performance of traditional GNNs in heterophilous settings and outperforms state-of-the-art models in most cases.

Ruifeng Li 0002, Mingqian Li, Wei Liu, Hongyang Chen 0001

Effective molecular representation learning is crucial for advancing molecular property prediction and drug design. Mainstream molecular representation learning methods rely on Graph Neural Networks (GNNs) to model complex molecular structures. However, these GNN-based approaches struggle with three significant issues: insufficient annotations, molecular diversity, and architectural limitations such as over-squashing, which leads to the loss of critical structural details. To address these issues, we introduce a new class of GNNs that integrates the Kolmogorov-Arnold Networks (KANs), known for their robust data-fitting capabilities and high accuracy in small-scale AI + Science tasks. By incorporating KANs into GNNs, our model enhances the representation of molecular structures. We further advance this approach with a variant called SwallowKAN (SKAN), which employs adaptive Radial Basis Functions (RBFs) as the core of the non-linear neurons. This innovation improves both computational efficiency and adaptability to diverse molecular structures. Building on the strengths of SKAN, we propose a new class of GNNs, GNN-SKAN, and its augmented variant, GNN-SKAN+, which incorporates a SKAN-based classifier to further boost the performance. Experiments across 6 classification datasets, 6 regression datasets, and 4 few-shot learning datasets demonstrate that our approach achieves new state-of-the-art performance in terms of accuracy and computational cost.

Guoming Li, Jian Yang 0035, Yifan Chen 0004

Filtering-based graph neural networks (GNNs) constitute a distinct class of GNNs that employ graph filters to handle graph-structured data, achieving notable success in various graph-related tasks. Conventional methods adopt a graph-wise filtering paradigm, imposing a uniform filter across all nodes, yet recent findings suggest that this rigid paradigm struggles with heterophilic graphs. To overcome this, recent works have introduced node-wise filtering, which assigns distinct filters to individual nodes, offering enhanced adaptability. However, a fundamental gap remains: a comprehensive framework unifying these two strategies is still absent, limiting theoretical insights into the filtering paradigms. Moreover, through the lens of Contextual Stochastic Block Model, we reveal that a synthesis of graph-wise and node-wise filtering provides a sufficient solution for classification on graphs exhibiting both homophily and heterophily, suggesting the risk of excessive parameterization and potential overfitting with node-wise filtering. To address the limitations, this paper introduces Coarsening-guided Partition-wise Filtering (CPF). CPF innovates by performing filtering on node partitions. The method begins with structure-aware partition-wise filtering, which filters node partitions obtained via graph coarsening algorithms, and then performs feature-aware partition-wise filtering, refining node embeddings via filtering on clusters produced by k-means clustering over features. In-depth analysis is conducted for each phase of CPF, showing its superiority over other paradigms. Finally, benchmark node classification experiments, along with a real-world graph anomaly detection application, validate CPF's efficacy and practical utility. Code is available with the Github repository: https://github.com/vasile-paskardlgm/CPF.

Gaotang Li, Danai Koutra, Yujun Yan

We address the key challenge of size-induced distribution shifts in graph neural networks (GNNs) and their impact on the generalization of GNNs to larger graphs. Existing literature operates under diverse assumptions about distribution shifts, resulting in varying conclusions about the generalizability of GNNs. In contrast to prior work, we adopt a data-driven approach to identify and characterize the types of size-induced distribution shifts and explore their impact on GNN performance from a spectral standpoint, a perspective that has been largely underexplored. Leveraging the significant variance in graph sizes in real biological datasets, we analyze biological graphs and find that spectral differences-driven by subgraph patterns (e.g., average cycle length)-strongly correlate with GNN performance on larger, unseen graphs. Based on these insights, we propose three model-agnostic strategies to enhance GNNs' awareness of critical subgraph patterns, identifying size-intensive attention as the most effective approach. Extensive experiments with six GNN architectures and seven model-agnostic strategies across five datasets show that our size-intensive attention strategy significantly improves graph classification on test graphs 2 to 10 times larger than the training graphs, boosting F1 scores by up to 8% over strong baselines.