WWW 2025 Papers

Processing long contexts presents a significant challenge for large language models (LLMs). While recent advancements allow LLMs to handle much longer contexts than before (e.g., 32K or 128K tokens), it is computationally expensive and can still be insufficient for many applications. Retrieval-Augmented Generation (RAG) is considered as a promising strategy to address this problem. However, conventional RAG methods face inherent limitations because of two underlying requirements: 1) explicitly-stated queries, and 2) well-structured knowledge. These conditions, however, do not hold in general long-context processing tasks. In this work, we propose HawkRAG, a novel RAG framework empowered by global memory-augmented retrieval. HawkRAG features a dual-system architecture. The name HawkRAG is inspired by the way a hawk glides high in the sky to observe the land, allowing it to spot and target prey with precision from a broad vantage point. First, it employs a light but long-range system to create a global memory of the long context. Once a task is presented, it generates draft answers, providing useful clues for the retrieval tools to locate relevant information within the long context. Second, it leverages an expensive but expressive system, which generates the final answer based on the retrieved information. Building upon this fundamental framework, we realize the memory module in the form of KV compression, and reinforce its memorization and cluing capacity from the Generation quality's Feedback (a.k.a. RLGF). In our experiments, HawkRAG achieves superior performances across a variety of long-context evaluation tasks, not only complex scenarios where traditional RAG methods struggle, but also simpler ones where RAG is typically applied.

The enhance of accuracy in reasoning results of LLMs arouses the community’s interests, wherein pioneering studies investigate post-hoc strategies to rectify potential mistakes. Despite extensive efforts, they are all stuck in a state of resource competition demanding significant time and computing expenses. The cause of the situation lies in failing to identify the fundamental feature of the solutions in this line, coined as the self-denial of LLMs. In other words, LLMs should confidently determine the potential mistakes and carefully execute the targeted correction. As the whole procedure conducts within LLMs, supporting and persuasive references are hard to acquire, while the absence of specific steps towards refining mistakes persists even when errors are acknowledged. In response to the challenges, we present PSSD, which refers to and implements the human psyche structure such that three distinct and interconnected roles contribute to human reasoning. Specifically, PSSD leverages the recent multi-agent paradigm, and is further enhanced with three innovatively conceived roles: (1) the intuition-based id role that provides initial attempts based on benign LLMs; (2) the rule-driven superego role that summarizes rules to regulate the above attempts, and returns specific key points as guidance; and (3) the script-centric ego role that absorbs all procedural information to generate executable script for the final answer prediction. Extensive experiments demonstrate that the proposed design not only better enhance reasoning capabilities, but also seamlessly integrate with current models, leading to superior performance.

Graphs are widely used to model complex systems across various domains, including social networks and biological systems. A key task in graph analysis is identifying recurring structural patterns, known as graphlets, which capture connectivity among a fixed-size subset of nodes. While graphlets have been extensively applied in tasks such as measuring graph similarity and identifying communities, conventional graphlets focus only on direct connections between nodes. This limitation overlooks potential insights from more distant relationships within the graph structure. In this paper, we introduce (𝑑, 𝑠)-graphlets, a generalization of size-𝑠 graphlets that incorporates indirect connections between nodes up to distance 𝑑. This new formulation provides a more fine-grained and comprehensive understanding of local graph structures. To efficiently count (𝑑, 𝑠)-graphlets in a graph, we present EDGE, an exact counting algorithm that employs optimized combinatorial techniques to significantly reduce computational complexity compared to naive enumeration. Our empirical analysis across diverse real-world datasets demonstrates that (𝑑, 𝑠)-graphlets provide superior graph characterization, outperforming conventional graphlets in the graph clustering task. Moreover, our case studies show that (𝑑, 𝑠)-graphlets uncover non-trivial insights that would remain undiscovered when using conventional graphlets.

With the explosive growth of online service platforms, an increasing number of people and enterprises are undertaking personal and professional tasks online. In real applications such as trip planning and online marketing, planning sessions for a sequence of activities or services will enable social users to receive the optimal services, improving their experience and reducing the cost of their activities. These online platforms are heterogeneous, including different types of services with different attributes. However, the problem of session planning over heterogeneous platforms has not been studied so far. In this paper, we propose a Motivation-Aware Session Planning (MASP) framework for session planning over heterogeneous social platforms. Specifically, we first propose a novel HeterBERT model to handle the heterogeneity of items at both type and attribute levels. Then, we propose to predict user preference using the motivations behind user activities. Finally, we propose an algorithm together with its optimisations for efficient session generation. The extensive tests prove the high effectiveness and efficiency of MASP.

Coordinated information operations remain a persistent challenge on social media, despite platform efforts to curb them. While previous research has primarily focused on identifying these operations within individual platforms, this study shows that coordination frequently transcends platform boundaries. Leveraging newly collected data of online conversations related to the 2024 U.S. Election across $\mathbb{X}$ (formerly Twitter), Facebook, and Telegram, we construct similarity networks to detect coordinated communities exhibiting suspiciously similar sharing behaviors within and across platforms. Introducing an advanced coordination detection model, we reveal evidence of potential foreign interference, with Russian-affiliated media being systematically promoted across Telegram and $\mathbb{X}$. Our analysis also uncovers substantial intra- and cross-platform coordinated inauthentic activity, driving the spread of highly partisan, low-credibility, and conspiratorial content. These findings highlight the urgent need for regulatory measures that extend beyond individual platforms to effectively address the growing challenge of cross-platform coordinated influence campaigns.

We explored in this work the ubiquitous phenomenon of serial scammers, who deploy thousands of addresses to conduct a series of similar Rug Pulls on popular decentralized exchanges (DEXs). We first constructed a list of about 163,000 scammer addresses behind all 1-day Rug Pulls on the two most popular DEXs, Uniswap (Ethereum) and Pancakeswap (BSC), and identified many distinctive scam patterns including star-shaped, chain-shaped and majority-flow scam clusters. We then proposed an algorithm to build a complete scam network from given scammer addresses, which consists of not only scammer addresses but also supporting addresses including depositors, withdrawers, transferrers, coordinators, and most importantly, wash traders. We note that profit estimations in existing works on Rug Pulls failed to capture the cost of wash trading, leading to inflated figures. Knowing who the wash traders are, we established a more accurate estimate for the true profit of individual scam pools as well as of the entire (serial) scam network by taking into account the wash-trading expenses.

Traffic prediction plays a pivotal role in intelligent transportation systems. Most existing studies only predict traffic flow for a specific time period based on traffic data from a short period, such as an hour, overlooking the influence of periodicity present in traffic data. Moreover, most of the existing advanced methods rely on manually constructed spatio-temporal graphs for joint modeling, or use pure spatial and pure temporal modules to separately model spatial and temporal features, which limits the learning of complex spatio-temporal patterns in traffic data due to structural inadequacies in the model. To address these issues, we propose a novel approach by constructing a learnable long-range spatio-temporal graph, which can better capture complex patterns in traffic data. We introduce a new model, LLGformer, which improves upon traditional Transformer-style models, facilitating more efficient learning of traffic flow data by integrating long-range historical information. Leveraging attention mechanisms on a spatiotemporal graph enables direct interaction of information across different time slices and locations. Additionally, we propose two optimization strategies to further boost the speed of training and inference. Extensive experiments on four real-world datasets show that the new model significantly outperforms state-of-the-art methods.

Machine unlearning is an emerging field that selectively removes specific data samples from a trained model. This capability is crucial for addressing privacy concerns, complying with data protection regulations, and correcting errors or biases introduced by certain data. Unlike traditional machine learning, where models are typically static once trained, machine unlearning facilitates dynamic updates that enable the model to "forget" information without requiring complete retraining from scratch. There are various machine unlearning methods, some of which are more time-efficient when data removal requests are fewer. To decrease the execution time of such machine unlearning methods, we aim to reduce the size of data removal requests based on the fundamental assumption that the removal of certain data would not result in a distinguishable retrained model. We first propose the concept of unnecessary unlearning, which indicates that the model would not alter noticeably after removing some data points. Subsequently, we review existing solutions that can be used to solve our problem. We highlight their limitations in adaptability to different unlearning scenarios and their reliance on manually selected parameters. We consequently put forward FUNU, a method to identify data points that lead to unnecessary unlearning. FUNU circumvents the limitations of existing solutions. The idea is to discover data points within the removal requests that have similar neighbors in the remaining dataset. We utilize a reference model to set parameters for finding neighbors, inspired from the area of model memorization. We provide a theoretical analysis of the privacy guarantee offered by FUNU and conduct extensive experiments to validate its efficacy.

Advanced Persistent Threats (APTs) are continuously evolving, leveraging their stealthiness and persistence to put increasing pressure on current provenance-based Intrusion Detection Systems (IDS). This evolution exposes several critical issues: (1) The dense interaction between malicious and benign nodes within provenance graphs introduces neighbor noise, hindering effective detection; (2) The complex prediction mechanisms of existing APTs detection models lead to the insufficient utilization of prior knowledge embedded in the data; (3) The high computational cost makes detection impractical. To address these challenges, we propose Vodka, a lightweight threat detection system built on a knowledge distillation framework, capable of node-level detection within audit log provenance graphs. Specifically, Vodka applies graph Laplacian regularization to reduce neighbor noise, obtaining smoothed and denoised graph signals. Subsequently, Vodka employs a teacher model based on GNNs to extract knowledge, which is then distilled into a lightweight student model. The student model is designed as a trainable combination of a feature transformation module and a personalized PageRank random walk label propagation module, with the former capturing feature knowledge and the latter learning label and structural knowledge. After distillation, the student model benefits from the knowledge of the teacher model to perform precise threat detection. Finally, Vodka reconstructs attack paths from anomalous nodes, providing insight into the attackers' strategies. We evaluate Vodka through extensive experiments on three public datasets and compare its performance against several state-of-the-art IDS solutions. The results demonstrate that Vodka achieves outstanding detection accuracy across all scenarios and the detection time is 1.4 to 5.2 times faster than the current state-of-the-art methods.

Retrieval-augmented generation (RAG) is a well-suited technique for retrieving privacy-sensitive Electronic Health Records (EHR). It can serve as a key module of the healthcare copilot, helping reduce misdiagnosis for healthcare practitioners and patients. However, the diagnostic accuracy and specificity of existing heuristic-based RAG models used in the medical domain are inadequate, particularly for diseases with similar manifestations. This paper proposes MedRAG, a RAG model enhanced by knowledge graph (KG)-elicited reasoning for the medical domain that retrieves diagnosis and treatment recommendations based on manifestations. MedRAG systematically constructs a comprehensive four-tier hierarchical diagnostic KG encompassing critical diagnostic differences of various diseases. These differences are dynamically integrated with similar EHRs retrieved from an EHR database, and reasoned within a large language model. This process enables more accurate and specific decision support, while also proactively providing follow-up questions to enhance personalized medical decision-making. MedRAG is evaluated on both a public dataset DDXPlus and a private chronic pain diagnostic dataset (CPDD) collected from our cooperating hospital, and its performance is compared against various existing RAG methods. Experimental results show that, leveraging the information integration and relational abilities of the KG, our MedRAG provides more specific diagnostic insights and outperforms state-of-the-art models in reducing misdiagnosis rates. Our code will be available at https://github.com/username00-c/MedRAG.git.

Climate activism is crucial in stimulating collective societal and behavioral change towards sustainable practices through political pressure. Although multiple factors contribute to the participation in activism, their complex relationships and the scarcity of data on their interactions have restricted most prior research to studying them in isolation, thus preventing the development of a quantitative, causal understanding of why people approach activism. In this work, we develop a comprehensive causal model of how and why Reddit users engage with activist communities driving mass climate protests (mainly the 2019 Earth Strike, Fridays for Future, and Extinction Rebellion). Our framework, based on Stochastic Variational Inference applied to Bayesian Networks, learns the causal pathways over multiple time periods. Distinct from previous studies, our approach uses large-scale and fine-grained longitudinal data (2016 to 2022) to jointly model the roles of sociodemographic makeup, experience of extreme weather events, exposure to climate-related news, and social influence through online interactions. We find that among users interested in climate change, participation in online activist communities is indeed influenced by direct interactions with activists and largely by recent exposure to media coverage of climate protests. Among people aware of climate change, left-leaning people from lower socioeconomic backgrounds are particularly represented in online activist groups. Our findings offer empirical validation for theories of media influence and critical mass, and lay the foundations to inform interventions and future studies to foster public participation in collective action.

Link prediction methods predict missing facts in incomplete knowledge graphs, often using embeddings to enhance scalability. However, embeddings complicate explainability, which is crucial for users' understanding of inferences in many domains. Methods emerged to explain predictions by identifying supporting portions of knowledge. To evaluate explanations from a user perspective, they can be compared to those in benchmarks, though they are limited to simplistic graphs. In contrast, user studies on forward simulatability variation measure how explanations improve predictability, i.e., the user ability to predict the results of inferences, which is key to trust. However, user studies face scalability and reproducibility issues on large graphs. Recognizing these gaps, we propose LP-DIXIT to algorithmically evaluate explanations of link predictions by determining forward simulatability variation and adopting large language models to mimic users, as is done in other domains, e.g., in evaluating other approaches on language related tasks. We experimentally prove that LP-DIXIT evaluates as effective explanations those in benchmarks, and we adopt it to compare state-of-the-art explanation methods.

Recently, many efforts have been devoted to building effective sequential recommenders. Despite their effectiveness, these methods typically develop a single model to serve all users. However, our empirical studies reveal that different sequential encoders have intrinsic architectural biases and tend to focus on specific behavioral patterns, \ie particular frequency range of user behavior sequences. For example, the Self-Attention module is essentially a low-pass filter, focusing on low-frequency information while neglecting the high-frequency details. This evidently limits their ability to capture diverse user patterns, leading to suboptimal recommendations. To tackle this problem, we present FamouSRec, a Frequency-Augmented mixture-of-Heterogeneous-Experts Framework for personalized recommendations. Our approach builds an MoE-based recommender system, integrating the strengths of various experts to achieve diversified user modeling. For developing the MoE framework, as the key to our approach, we instantiate experts with various model architectures, aiming to leverage their inherent architectural biases and capture diverse behavioral patterns. For selecting appropriate experts to serve individuals, we introduce a frequency-augmented router. It first identifies frequency components in user behavior sequences that are suited for expert encoding, and then conducts customized routing based on the informativeness of these components. Building on this framework, we further propose two novel contrastive tasks to enhance expert specialization and alignment, thus further improving modeling efficacy and enabling robust recommendations. Extensive experiments on five real-world datasets demonstrate the effectiveness of our approach.

Personalized learner modeling uses learners' historical behavior data to diagnose their cognitive abilities, a process known as Cognitive Diagnosis (CD) in the literature. This is a fundamental yet crucial task in web-based learning services, such as learning resource recommendation and adaptive testing. Previously, researchers discovered that models improperly correlate learners' abilities with their sensitive attributes, resulting in unfair diagnoses for learners from different sensitive groups (e.g., gender, region). Given the input of sensitive attributes, researchers proposed decorrelating these attributes from the modeling process, demonstrating improved fairness results. However, privacy concerns make collecting sensitive attributes impractical. This challenge is compounded by the presence of multiple sensitive attributes, making fairness improvement under any of them difficult. In this paper, we explore how to achieve fair personalized learner modeling without relying on any sensitive attribute input. Specifically, we first introduce a novel fairness objective tailored for personalized learner modeling. We then propose a max-min strategy that facilitates both potential sensitive information inference and fair CD modeling. In the max step, we propose a pseudo-label inference method based on maximizing the designed fairness objective. Given these pseudo-labels, the min step involves retraining a fair CD model by minimizing the designed objective. Additionally, we provide a theoretical guarantee that implementing our proposed framework reduces the upper bound of fairness generalization error. Extensive experiments demonstrate that the proposed framework significantly outperforms existing methods in terms of fairness and accuracy. Our code is available at https://anonymous.4open.science/r/FairWISA-40C6/.

Graph self-supervised learning has gained significant attention recently. However, many existing approaches heavily depend on perturbations, and inappropriate perturbations may corrupt the graph’s inherent information. The Vector Quantized Variational Autoencoder (VQ-VAE) is a powerful autoencoder extensively used in fields such as computer vision; however, its application to graph data remains underexplored. In this paper, we provide an empirical analysis of vector quantization in the context of graph autoencoders, demonstrating its significant enhancement of the model's capacity to capture graph topology. Furthermore, we identify two key challenges associated with vector quantization when applying in graph data: codebook underutilization and codebook space sparsity. For the first challenge, we propose an annealing-based encoding strategy that promotes broad code utilization in the early stages of training, gradually shifting focus toward the most effective codes as training progresses. For the second challenge, we introduce a hierarchical two-layer codebook that captures relationships between embeddings through clustering. The second layer codebook links similar codes, encouraging the model to learn closer embeddings for nodes with similar features and structural topology in the graph. Our proposed model outperforms 16 representative baseline methods in self-supervised link prediction and node classification tasks across multiple datasets. Our implementation is available at https://anonymous.4open.science/r/hqa-gae-D2F4.

With the development of multi-modality data modeling techniques, recent recommender systems use not only textual data and user-item interactions but also multi-modality data such as images to improve their performances. Existing methods typically adopt cross-modal pairwise alignment strategies to alleviate the gap between modalities. Nevertheless, this alignment paradigm has limitations on explainability, consistency, and expansibility, which may only achieve suboptimal performances. In this paper, we propose a novel Explainable generative multi-modality Alignment method for transferable Recommender systems, i.e., EARec. Specifically, we design a two-stage pipeline to achieve unified multi-modality alignment of items and the sequential recommendation task, respectively. In the first phase, we present a generation task that parallel aligns each modality from multiple source domains to an anchor with explainable meaning. Three modality features share the same anchor to achieve a consistent alignment direction. Additionally, we incorporate behavior-related information as an independent modality into the alignment framework, establishing a bridge that promotes the alignment between multi-modalities and behavior. In the second stage, we composite the aligned modality encoders into a unified one and then transfer it to the target domain to enhance sequential recommendation. The pipeline that adopts parallel multi-modal alignment and composition shows flexibility and scalability for incorporating new modalities. Experimental results on multiple public datasets demonstrate the superiority of EARec over multi-modality recommendation baselines and further analysis indicates the explainability of generative alignment.

Platforms design the forms of presentation by which sellers are shown to the buyers. This design not only shapes the buyers' experience but also leads to different market equilibria or dynamics. One component in this design is through the platform's mediation of the search frictions experienced by the buyers for different sellers. We take a model of monopolistic competition and show that, on one hand, when all sellers have the same inspection costs, the market sees no stable price as the sellers always have incentive to undercut each other, and, on the other hand, the platform may stabilize the price by giving prominence to one seller chosen by a carefully designed mechanism. This calls to mind Amazon's Buy Box design. We study natural mechanisms for choosing the prominent seller, characterize the range of equilibrium prices implementable by them, and find, somewhat counterintuitively, that in certain scenarios the buyers' surplus improves as the search friction increases.

Graph prompt tuning has achieved significant success for its ability to effectively adapt pre-trained graph neural networks to various downstream tasks. However, the pre-trained models may learn discriminatory representation due to the inherent prejudice in graph-structured data. Existing graph prompt tuning overlooks such unfairness, leading to biased outputs towards certain demographic groups determined by sensitive attributes such as gender, age, and political ideology. To overcome this limitation, we propose a fairness-aware graph prompt tuning method to promote fairness while enhancing the generality of any pre-trained GNNs (named FPrompt). FPrompt introduces hybrid graph prompts to augment counterfactual data while aligning the pre-training and downstream tasks. It also applies edge modification to increase sensitivity heterophily. We provide a two-fold theoretical analysis: first, we demonstrate that FPrompt possesses universal capabilities in handling pre-trained GNN models across various pre-training strategies, ensuring its adaptability in different scenarios. Second, we show that FPrompt effectively reduces the upper bound of generalized statistical parity, thereby mitigating the bias of pre-trained models. Extensive experiments demonstrate that FPrompt outperforms baseline models in both accuracy and fairness (~$33$\%) on benchmark datasets. Additionally, we introduce a new benchmark for transferable evaluation, showing that FPrompt achieves state-of-the-art generalization performance.

The field of Reinforcement Learning (RL) has garnered increasing attention for its ability of optimizing user retention in recommender systems. A primary obstacle in this optimization process is the environment non-stationarity stemming from the continual and complex evolution of user behavior patterns over time, such as variations in interaction rates and retention propensities. These changes pose significant challenges to existing RL algorithms for recommendations, leading to issues with dynamics and reward distribution shifts. This paper introduces a novel approach called Adaptive User Retention Optimization (AURO) to address this challenge. To navigate the recommendation policy in non-stationary environments, AURO introduces an state abstraction module in the policy network. The module is trained with a new value-based loss function, aligning its output with the estimated performance of the current policy. As the policy performance of RL is sensitive to environment drifts, the loss function enables the state abstraction to be reflective of environment changes and notify the recommendation policy to adapt accordingly. Additionally, the non-stationarity of the environment introduces the problem of implicit cold start, where the recommendation policy continuously interacts with users displaying novel behavior patterns. AURO encourages exploration guarded by performance-based rejection sampling to maintain a stable recommendation quality in the cost-sensitive online environment. Extensive empirical analysis are conducted in a user retention simulator, the MovieLens dataset, and a live short-video recommendation platform, demonstrating AURO's superior performance against all evaluated baseline algorithms.