WWW 2025 论文列表

Money laundering is the process that intends to legalize the income derived from illicit activities, thus facilitating their entry into the monetary flow of the economy without jeopardizing their source. It is crucial to identify such activities accurately and reliably in order to enforce anti-money laundering (AML). Despite considerable efforts to AML, a large number of such activities still go undetected. Rule-based methods were first widely used in the early days and still be widely used in existing detection systems. With the rise of machine learning, graph-based learning methods have gained prominence in detecting illicit accounts by analyzing money transfer graphs between accounts. However, existing approaches work based on the prerequisite that the transaction graph is centralized, while in practice, money laundering activities usually span multiple financial institutions. Due to regulatory, legal, commercial, and customer privacy concerns, institutions tend not to share data, limiting their utility in practical usage. In this paper, we propose the first algorithm that supports performing AML over multiple institutions while protecting the security and privacy of local data. To evaluate, we construct Alipay-ECB, a real-world dataset comprising digital transactions from Alipay, the world’s largest mobile payment platform, alongside transactions from E-Commerce Bank (ECB). The dataset includes over 200 million accounts and 300 million transactions, covering both intra-institution transactions and those between Alipay and ECB. This makes it the largest real-world transaction graph available for analysis. The experimental results demonstrate that our methods can effectively identify cross-institution money laundering subgroups. Additionally, experiments on synthetic datasets also demonstrate that our method is efficient, requiring only a few minutes on datasets with millions of transactions.

Hierarchical text classification aims to categorize each document into a set of classes in a label taxonomy, which is a fundamental web text mining task with broad applications such as web content analysis and semantic indexing. Most earlier works focus on fully or semi-supervised methods that require a large amount of human annotated data which is costly and time-consuming to acquire. To alleviate human efforts, in this paper, we work on hierarchical text classification with a minimal amount of supervision: using the sole class name of each node as the only supervision. Recently, large language models (LLM) show competitive performance on various tasks through zero-shot prompting, but this method performs poorly in the hierarchical setting because it is ineffective to include the large and structured label space in a prompt. On the other hand, previous weakly-supervised hierarchical text classification methods only utilize the raw taxonomy skeleton and ignore the rich information hidden in the text corpus that can serve as additional class-indicative features. To tackle the above challenges, we propose TELEClass, **T**axonomy **E**nrichment and **L**LM-**E**nhanced weakly-supervised hierarchical text **Class**ification, which combines the general knowledge of LLMs and task-specific features mined from an unlabeled corpus. TELEClass automatically enriches the raw taxonomy with class-indicative features for better label space understanding and utilizes novel LLM-based data annotation and generation methods specifically tailored for the hierarchical setting. Experiments show that TELEClass can significantly outperform previous strong baselines while also achieving comparable performance to zero-shot prompting of LLMs with drastically less inference cost.

Memes serve as a powerful medium of expression in the digital age, shaping cultural discourse and conveying ideas succinctly and engagingly. However, their potential for social abuse highlights the importance of developing effective methods to detect harmful content within memes. Recent studies on memes have focused on transforming images into textual captions using large language models (LLMs). However, these approaches often result in non-informative captions. Furthermore, previous methods have only been tested on limited datasets, providing insufficient evidence of their robustness. To address these limitations, we present a multimodal, agent-based framework designed to generate informative visual descriptions of memes by asking insightful questions to improve visual descriptions in zero-shot visual question-answering settings. Specifically, we leverage an LLM as agents with distinct roles and a large multimodal model (LMM) as a vision expert. These agents first analyze the images and then ask informative questions related to potential social abuse in memes to obtain high-quality answers about the images. Through continuous discussion guided by instructional prompts, the agents gather high-quality information while repeatedly acquiring image data from the LMM, which helps detect social abuse in memes. Finally, the discussion history and basic information are classified using the LLM to obtain the final prediction results in a zero-shot setting. Experimental results on a meme benchmark dataset sourced from 5 diverse meme datasets, comprising 6,626 memes spanning 5 tasks of varying complexity related to social abuse, demonstrate that our framework outperforms state-of-the-art methods, with detailed comparative analysis and ablation studies, further validating its generalizability and ability to retrieve more relevant information for detecting social abuse in memes.

CAPTCHAs have long been essential tools for protecting applications from automated bots. Initially designed as simple questions to distinguish humans from bots, they have become increasingly complex to keep pace with the proliferation of CAPTCHA-cracking techniques employed by malicious actors. However, with the advent of advanced large language models (LLMs), the effectiveness of existing CAPTCHAs is now being undermined. To address this issue, we have conducted an empirical study to evaluate the performance of multimodal LLMs in solving CAPTCHAs and to assess how many attempts human users typically need to pass them. Our findings reveal that while LLMs can solve most CAPTCHAs, they struggle with those requiring complex reasoning—a type of CAPTCHA that also presents significant challenges for human users. Interestingly, our user study showed that the majority of human participants required a second attempt to pass these reasoning CAPTCHAs, a finding not previously reported in existing research. Based on the findings of our empirical study, we introduce IllusionCAPTCHA, an innovative approach designed to be "Human-Easy but AI-Hard". This new CAPTCHA employs visual illusions to create tasks that are intuitive for humans but highly confusing for AI models. Furthermore, we developed a structured, step-by-step method that guides LLMs toward making specific incorrect choices, thereby reducing their ability to bypass CAPTCHA systems successfully. Our evaluation shows that IllusionCAPTCHA can effectively deceive LLMs 100\% of the time. Moreover, our structured design significantly increases the likelihood of AI errors when attempting to solve these challenges. Results from our user study indicate that 86.95\% of participants successfully passed the CAPTCHA on their first attempt, outperforming other CAPTCHA systems.

Large language models (LLMs) demonstrate significant potential in various applications; however, they are susceptible to generating hallucinations, which can lead to the spread of misinformation online. Existing studies address hallucination detection by (1) employing reference-based methods that consult external resources for verification or (2) utilizing reference-free methods that mainly estimate answer uncertainty based on LLM's internal states. However, reference-based methods incur significant costs and can be infeasible for obtaining reliable external references. Besides, existing uncertainty estimation (UE) methods often overlook the impact of scenario backgrounds inherited from the query's lexical resources, leading to noise in UE. In almost all real-world applications, users care about the uncertainty concerning semantics or facts instead of the query's scenario information. Therefore, we argue that mitigating scenario-related noise and focusing on semantic information can yield a more desirable UE. In this paper, we introduce a plug-and-play scenario-independent framework to enhance unsupervised UE in LLMs by removing scenario-related noise and focusing on semantic information. This framework is compatible with most existing UE methods, as it leverages only the existing UE methods' outputs. Specifically, we design a scenario-specific sampling to paraphrase queries, maintaining their common semantics while diversifying the scenario distribution. Subsequently, to estimate the contribution of the common semantics, we design a factor analysis (FA) model to disentangle the UE score obtained from the given UE method into a combination of multiple latent factors, which represent the contribution of the common semantics and scenario-related noise. By solving the FA model, we decompose the impact of the most significant factor to approximate the uncertainty caused by the common semantics, thus achieving scenario-independent UE. Extensive experiments and analysis across multiple models and datasets demonstrate the effectiveness of our approach.

Building location embedding from web-sourced satellite imagery has emerged as an enduring research focus in web mining. However, most existing methods are inherently constrained by their reliance on discrete, sparse sampling strategies, failing to capture the essential spatial continuity of geographic spaces. Moreover, the presence of confounding factors in satellite images can distort the perception of actual objects, leading to semantic discontinuity in the embeddings. In this work, we propose **SatCLE**, a novel framework for Continuous Location Embeddings leveraging Satellite imagery. Specifically, to address the out-of-distribution query challenge of spatial continuity, we propose a geospatial refinement strategy comprising stochastic perturbation continuity expansion and graph propagation fusion, which transforms discrete geospatial coordinates into a continuous space. To mitigate the effects of confounders on semantic continuity, we introduce causal refinement, integrating causal theory to localize and eliminate spurious correlations arising from the environmental context. Through extensive experiments, **SatCLE** shows state-of-the-art performance, exhibiting superior spatial coherence and semantic fidelity across diverse geospatial tasks.

Empathetic Response Generation (ERG) is one of the key tasks of the affective computing area, which aims to produce emotionally nuanced and compassionate responses to user's queries. However, existing ERG research is predominantly confined to the singleton text modality, limiting its effectiveness since human emotions are inherently conveyed through multiple modalities. To combat this, we introduce an avatar-based Multimodal ERG (MERG) task, entailing rich text, speech, and facial vision information. We first present a large-scale high-quality benchmark dataset, \textbf{AvaMERG}, which extends traditional text ERG by incorporating authentic human speech audio and dynamic talking-face avatar videos, encompassing a diverse range of avatar profiles and broadly covering various topics of real-world scenarios. Further, we deliberately tailor a system, named \textbf{Empatheia}, for MERG. Built upon a Multimodal Large Language Model (MLLM) with multimodal encoder, speech and avatar generators, Empatheia performs end-to-end MERG, with Chain-of-Empathetic reasoning mechanism integrated for enhanced empathy understanding and reasoning. Finally, we devise a list of empathetic-enhanced tuning strategies, strengthening the capabilities of emotional accuracy and content, avatar-profile consistency across modalities. Experimental results on AvaMERG data demonstrate that Empatheia consistently shows superior performance than baseline methods on both textual ERG and MERG. Overall, this work is expected to pioneer the MERG research by introducing a novel benchmark and an end-to-end model, laying a solid foundation for future advancements in multimodal empathetic response generation.

This paper proposes CrossLink, a novel framework for cross-domain link prediction. CrossLink learns the evolution pattern of a specific downstream graph and subsequently makes pattern-specific link predictions. It employs a technique called \textit{conditioned link generation}, which integrates both evolution and structure modeling to perform evolution-specific link prediction. This conditioned link generation is carried out by a transformer-decoder architecture, enabling efficient parallel training and inference. CrossLink is trained on extensive dynamic graphs across diverse domains, encompassing 6 million dynamic edges. Extensive experiments on eight untrained graphs demonstrate that CrossLink achieves state-of-the-art performance in cross-domain link prediction. Compared to advanced baselines under the same settings, CrossLink shows an average improvement of 11.40% in Average Precision across eight graphs. Impressively, it surpasses the fully supervised performance of 8 advanced baselines on 6 untrained graphs.

How can we efficiently identify the densest subgraph over relational graphs? Existing dense subgraph discovery (DSD) approaches assume that a relational graph $H$ is already derived from a heterogeneous data source and they focus on efficient discovery of the densest subgraph on the materialized $H$. Unfortunately, materializing relational graphs can be resource-intensive, which thus limits the practical usefulness of existing algorithms over large datasets. To mitigate this, we propose a novel Summary-bAsed deNsest Subgraph discovery (SANS) system. Our unique summary-based peeling algorithm forms the core of SANS. Following the peeling paradigm, it utilizes summaries of each node's neighborhood to efficiently estimate peeling coefficients and subgraph densities at each peeling iteration and thus avoids materializing the relational graph completely. Through extensive experiments, we demonstrate the efficacy and efficiency of SANS, reaching orders of magnitude speedups compared to the conventional baselines with materialization, while consistently achieving at least 95% accuracy compared to peeling algorithms based on materialization.

Retrieval-augmented generation (RAG) has demonstrated effectiveness in mitigating the hallucination problem of large language models (LLMs). However, the difficulty of aligning the retriever with the diverse LLMs' knowledge preferences inevitably poses an inevitable challenge in developing a reliable RAG system. To address this issue, we propose DPA-RAG, a universal framework designed to align diverse knowledge preferences within RAG systems. Specifically, we initially introduce a preference knowledge construction pipline and incorporate five novel query augmentation strategies to alleviate preference data scarcity. Based on preference data, DPA-RAG accomplishes both external and internal preference alignment: 1) It jointly integrate pair-wise, point-wise, and contrastive preference alignment abilities into the reranker, achieving external preference alignment among RAG components. 2) It further introduces a pre-aligned stage before vanilla Supervised Fine-tuning (SFT), enabling LLMs to implicitly capture knowledge aligned with their reasoning preferences, achieving LLMs' internal alignment. Experimental results across four knowledge-intensive QA datasets demonstrate that DPA-RAG outperforms all baselines and seamlessly integrates both black-box and open-sourced LLM readers. Further qualitative analysis and discussions also provide empirical guidance for achieving reliable RAG systems.

Item Response Theory (IRT) has been widely used in educational psychometrics to assess student ability, as well as the difficulty and discrimination of test questions. In this context, discrimination specifically refers to how effectively a question distinguishes between students of different ability levels, and it does not carry any connotation related to fairness. In recent years, IRT has been successfully used to evaluate the predictive performance of Machine Learning (ML) models, but this paper marks its first application in fairness evaluation. In this paper, we propose a novel Fair-IRT framework to evaluate a set of predictive models on a set of individuals, while simultaneously eliciting specific parameters, namely, the ability to make fair predictions (a feature of predictive models), as well as the discrimination and difficulty of individuals that affect the prediction results. Furthermore, we conduct a series of experiments to comprehensively understand the implications of these parameters for fairness evaluation. Detailed explanations for item characteristic curves (ICCs) are provided for particular individuals. We propose the flatness of ICCs to disentangle the unfairness between individuals and predictive models. The experiments demonstrate the effectiveness of this framework as a fairness evaluation tool. Two real-world case studies illustrate its potential application in evaluating fairness in both classification and regression tasks. Our paper aligns well with the Responsible Web track by proposing a Fair-IRT framework to evaluate fairness in ML models, which directly contributes to the development of a more inclusive, equitable, and trustworthy AI.

Dynamic Dependence Network (DDN) inference is crucial for understanding evolving relationships in multimodal time series web data, with broad applications in fields like medical and financial network analysis. The inherent dynamic nature, temporal continuity, and heterogeneous data sources in multimodal time series data pose three fundamental challenges: computational efficiency, prediction stability and robustness, and modality quality disparity. Previous methods, generally lacking utilization of multiple modalities, either struggle with computational efficiency due to the time-intensive manual hyperparameter tuning, or compromise prediction stability and robustness by neglecting temporal coherence. To address these challenges, we propose a Normalized mutual information-driven Tuning-free Dynamic Dependence Network inference method for multimodal data, namely NoTeNet. NoTeNet provides a promising paradigm that can integrate two different data modalities to enhance prediction accuracy. It uses normalized mutual information transforms noisy auxiliary data into relationship matrices and employs a kernel function for smooth temporal estimation. Additionally, NoTeNet significantly reduces the need for manual hyperparameter adjustments, offering a tuning-free approach with theoretical guarantees. On various synthetic datasets and real-world data, NoTeNet demonstrates superior prediction accuracy and efficiency without the need for hyperparameter tuning, making it potential for a wide range of web data applications.

We study the machine unlearning problem which aims to remove specific training data from a pre-trained machine learning model to allow users to exercise their ‘right to be forgotten’ to protect user privacy. Conventional machine unlearning methods would degrade the model performance after the unlearning procedure. To mitigate the issue, they typically rely on the access to the remaining training data to fine-tune the unlearned model to mitigate the influence of unlearning. However, accessing the remaining training data may not always be practical for different reasons (e.g., data expiration policies, storage limitations, or additional privacy constraints). Machine unlearning without access to the remaining training data poses significant challenges to retaining model performance. In this paper, we study how to unlearn specific training data from a pre-trained model without accessing the remaining training data and protect model performance without dramatically changing the model's parameters. We propose a practical method called Targeted Label Noise Injection. Intuitively, our method assigns incorrect yet controllable labels to the examples that need to be forgotten and fine-tunes the pre-trained model to learn these new labels. This strategy effectively moves the to-be-forgotten examples across the decision boundary with a small impact on the model's overall performance. We theoretically prove the effectiveness of the proposed method and empirically show that it achieves state-of-the-art unlearning performance across various datasets.