WWW 2025 Papers

Multi-modal sequential recommendation (SR) leverages multi-modal data to learn more comprehensive item features and user preferences than traditional SR methods, which has become a critical topic in both academia and industry. Existing methods typically focus on enhancing multi-modal information utility through adaptive modality fusion to capture the evolving of user preference from user-item interaction sequences. However, most of them overlook the interference caused by redundant interest-irrelevant information contained in rich multi-modal data. Additionally, they primarily rely on implicit temporal information based solely on chronological ordering, neglecting explicit temporal signals that could more effectively represent dynamic user interest over time. To address these limitations, we propose a Hierarchical time-aware Mixture of experts for multi-modal Sequential Recommendation (HM4SR) with a two-level Mixture of Experts (MoE) and a multi-task learning strategy. Specifically, the first MoE, named Interactive MoE, extracts essential user interest-related information from the multi-modal data of each item. Then, the second MoE, termed Temporal MoE, captures user dynamic interests by introducing explicit temporal embeddings from timestamps in modality encoding. To further address data sparsity, we propose three auxiliary supervision tasks: sequence-level category prediction (CP) for item feature understanding, contrastive learning on ID (IDCL) to align sequence context with user interests, and placeholder contrastive learning (PCL) to integrate temporal information with modalities for dynamic interest modeling. Extensive experiments on four public datasets verify the effectiveness of HM4SR compared to several state-of-the-art approaches.

Multimodal information extraction on social media is a series of fundamental tasks to construct the multimodal knowledge graph. The tasks are defined to extract the structural information in free texts with the incorporate images, such as: multimodal named entity typing and multimodal relation extraction. However, the growing number of multimodal data implies a growing category set and the newly emerged entity types or relations should be recognized without additional training. To address the aforementioned disadvantages, we focus on the zero-shot multimodal information extraction task which requires to utilize textual and visual modalities for identifying previously unseen categories in a zero-shot manner. Compared with the text-based zero-shot information extraction models, the existing multimodal ones make the textual and visual modalities aligned directly and exploit various fusion strategies to improve their generalization ability. But the existing methods only align the global representations of multimodal data and ignore the fine-grained semantic correlation of the text-image pairs and samples. Therefore, we propose the multimodal graph-based variational mixture of experts network (MG-VMoE) which takes the MoE network as the backbone and exploits the sparse expert weights for aligning the multimodal representations in a fine-grained way. Considering to learn the informative and aligned representations of multimodal data, we design each expert network as a variational information bottleneck to process the two modalities in a uni-backbone. Moreover, we do not only model the correlation of the text-image pair inner a sample, but also propose the multimodal graph-based virtual adversarial training to learn the semantic correlation between the samples. The experimental results on the two benchmark datasets demonstrate the superiority of MG-VMoE over the baselines.

Multi-modal time series data is common in web technologies like the Internet of Things (IoT). Existing methods for multi-modal time series representation learning aim to disentangle the modality-shared and modality-specific latent variables. Although achieving notable performances on downstream tasks, they usually assume an orthogonal latent space. However, the modality-specific and modality-shared latent variables might be dependent on real-world scenarios. Therefore, we propose a general generation process, where the modality-shared and modality-specific latent variables are dependent, and further develop a \textbf{M}ulti-mod\textbf{A}l \textbf{TE}mporal Disentanglement (\textbf{MATE}) model. Specifically, our \textbf{MATE} model is built on a temporally variational inference architecture with the modality-shared and modality-specific prior networks for the disentanglement of latent variables. Furthermore, we establish identifiability results to show that the extracted representation is disentangled. More specifically, we first achieve the subspace identifiability for modality-shared and modality-specific latent variables by leveraging the pairing of multi-modal data. Then we establish the component-wise identifiability of modality-specific latent variables by employing sufficient changes of historical latent variables. Extensive experimental studies on 12 datasets show a general improvement in different downstream tasks, highlighting the effectiveness of our method in real-world scenarios.

It is essential for e-commerce platforms to provide accurate, complete, and timely product attribute values, in order to improve the search and recommendation experience for both customers and sellers. In the real-world scenario, it is difficult for these platforms to identify attribute values for the newly introduced products given no similar product history records for training or retrieval. Besides, how to jointly learn the product representation given various product information in multiple modalities, such as textual modality (e.g., product titles and descriptions) and visual modality (e.g., product images), is also a challenging task. To address these limitations, we propose a novel method for extracting multi-label product attribute-value pairs from multiple modalities in the zero-shot scenario, where labeled data is absent during training. Specifically, our method constructs heterogeneous hypergraphs, where product information from different modalities is represented by different types of nodes, and the text and image nodes are embedded and learned through CLIP encoders to effectively capture and integrate multimodal product information. Then, the complex interrelations among these nodes are modeled through the hyperedges. By learning informative node representations, our method can accurately predict links between unseen product nodes and attribute-value nodes, enabling zero-shot attribute value extraction. We conduct extensive experiments and ablation studies on several categories of the public MAVE dataset and the results demonstrate that our proposed method significantly outperforms several state-of-the-art generative model baselines in multi-label, multi-modal product attribute value extraction in the zero-shot setting.

Federated Learning (FL) has emerged as a promising training framework that enables a server to effectively train a global model by coordinating multiple devices, i.e., clients, without sharing their raw data. Keeping data locally can ensure data privacy, but also makes the server difficult to assess data quality, leading to the noisy data issue. Specifically, for any given taring task, only a portion of each client's data is relevant and beneficial, while the rest may be redundant or noisy. Training with excessive noisy data can degrade performance. Motivated by this, we investigate the limitations of existing studies and develop an incentive mechanism with flexible pricing tailored for noisy data settings. The insight lies in mitigating the impact of noisy data by selecting appropriate clients and incentivizing them to clean their data spontaneously. Further, both rigorous theoretical analysis and extensive simulations compared with state-of-the-art methods have been well-conducted to validate the effectiveness of the proposed mechanism.

As a prevalent paradigm of modern web applications, cloud computing has experienced a surge in adoption. The deployment of vast and various workloads encapsulated within containers has become ubiquitous across cloud platforms, imposing substantial demands on the supporting infrastructure. However, Kubernetes (k8s), the de-facto standard for container orchestration, struggles with low scheduling throughput and high latency in large-scale clusters. The primary challenges are identified as excessive loads of read requests and resource contention among co-located components. In response to these challenges, in this paper, we present MerKury, a lightweight framework to enhance the Kubernetes performance for large-scale clusters. It employs a dual strategy: first, it preprocesses specific requests to alleviate unnecessary load, and second, it introduces an adaptive resource allocation algorithm to mitigate resource contention. Evaluations under different scenarios of varying cluster scale have demonstrated that MerKury notably augments cluster scheduling throughput up to 16.4$\times$ and reduces request latency by up to 39.3\%, outperforming vanilla Kubernetes and baseline resource allocation methods.

In sequential recommendation scenarios, user intent is a key driver of consumption behavior. However, consumption intents are usually latent and hence, difficult to leverage for recommender systems. Additionally, intents can be of repeated nature (e.g. yearly shopping for christmas gifts or buying a new phone), which has not been exploited by previous approaches. To navigate these impediments we propose the HyperHawkes framework which models user sessions via hypergraphs and extracts user intents via contrastive clustering. We use Hawkes Processes to model the temporal dynamics of intents, namely repeated consumption patterns and long-term interests of users. For short-term interest adaption, which is more fine-grained than intent-level modeling, we use a multi-level attention mixture network and fuse long-term and short-term signals. We use the generalized expectation-maximization (EM) framework for training the model by alternating between intent representation learning and optimizing parameters of the long- and short-term modules. Extensive experiments on four real-world datasets from different domains show that HyperHawkes significantly outperforms existing state-of-the-art methods.

Multi-modal recommender systems (MMRecs) leverage diverse modalities to deliver personalized recommendations, yet they often struggle with efficiency due to the large size of modality encoders and the complexity of fusing high-dimensional features. To address the efficiency issue, a promising solution is to compress a cumbersome MMRec into a lightweight ID-based Multi-Layer Perceptron-based Recommender system (MLPRec) through Knowledge Distillation (KD). Despite effectiveness, this approach overlooks the significant gap between the complex teacher MMRec and the lightweight, ID-based student MLPRec, which differ significantly in size, architecture, and input modalities, leading to ineffective knowledge transfer and suboptimal student performance. To bridge this gap, we propose TARec, a novel teacher-assisted Wasserstein Knowledge Distillation framework for compressing MMRecs into an efficient MLPRec. TARec introduces: (i) a two-staged KD process using an intermediate Teacher Assistant (TA) model to bridge the gap between teacher and student, facilitating smoother knowledge transfer; (ii) logit-level KD using the Wasserstein Distance as metric, replacing the conventional KL divergence to ensure stable gradient flow even with significant teacher-student gaps; and (iii) embedding-level contrastive KD to further distill high-quality embedding-level knowledge from teacher. Extensive experiments on real-world datasets verify the effectiveness of TARec, demonstrating that TARec significantly outperforms the state-of-the-art MMRecs while reducing computational costs. Our anonymous code is available at: https://anonymous.4open.science/r/TARec-0980/.

In digital education, Cognitive Diagnosis (CD) is essential for modeling learners' cognitive states, such as problem-solving ability and knowledge proficiency, by analyzing their response data, like answer correctness. However, traditional CD methods struggle with \textit{effectiveness} and \textit{efficiency}. They fail to capture the diversity and uncertainty of learners' cognitive states. Additionally, response prediction can be time-consuming. To address these issues, we propose BoxCD, a contrastive probabilistic box embedding model for cognitive diagnosis. BoxCD utilizes high-dimensional axis-aligned hyper-rectangles (boxes) to represent learners and exercises, with the volume of intersecting boxes used to predict learners' responses. This approach effectively captures semantic diversity and uncertainty while enhancing diagnostic effectiveness. To stabilize box embeddings, we integrate contrastive learning objectives with response prediction goals, optimizing the distance between positive and negative samples of learner and exercise boxes to improve uniformity. Additionally, we develop a rank-based response prediction method that leverages the geometric properties of box embeddings to efficiently assess learners' response correctness. Comprehensive experiments on two real-world datasets demonstrate that BoxCD outperforms traditional CD models in both effectiveness and efficiency, showcasing its potential to enhance personalized learning in digital education platforms.

Deep learning-based (DL-based) malicious traffic detection methods are effective but vulnerable to adversarial attacks. Existing adversarial attack methods have shown promising results when targeting traffic detection models based on statistics and sequence features. However, these methods are less effective against models that rely on payload analysis. The main reason is the difficulty in generating semantic, compliant, and functional payloads, which limits their practical application. In this paper, we propose AdvTG, an adversarial traffic generation framework based on the large language model (LLM) and reinforcement learning (RL). Specifically, AdvTG is designed to attack various DL-based detection models across diverse features and architectures, thereby enhancing the generalization capabilities of the generated adversarial traffic. Moreover, we design a specialized prompt for payload generation tasks, where functional fields and target types are supplied as input, while non-functional fields are generated to produce the mutated traffic. This fine-tuning endows the LLM with task comprehension and traffic pattern reasoning abilities, allowing it to generate traffic that remains compliant and functional. Furthermore, leveraging RL, AdvTG automatically selects traffic fields that exhibit more robust adversarial properties. Experimental results show that AdvTG achieves over 40\% attack success rate (ASR) across six detection models on four base datasets and two extended datasets, significantly outperforming other adversarial attack methods.

Data summarization tasks are often modeled as $k$-clustering problems, where the goal is to choose $k$ data points, called cluster centers, that best represent the dataset by minimizing a clustering objective. A popular objective is to minimize the maximum distance between any data point and its nearest center, which is formalized as the $k$-center problem. While in some applications all data points can be chosen as centers, in the general setting, centers must be chosen from a predefined subset of points, referred as facilities or suppliers; this is known as the $k$-supplier problem. In this work, we focus on fair data summarization modeled as the fair $k$-supplier problem, where data consists of several groups, and a minimum number of centers must be selected from each group while minimizing the $k$-supplier objective. The groups can be disjoint or overlapping, leading to two distinct problem variants each with different computational complexity. We present $3$-approximation algorithms for both variants, improving the previously known factor of $5$. For disjoint groups, our algorithm runs in polynomial time, while for overlapping groups, we present a fixed-parameter tractable algorithm, where the exponential runtime depends only on the number of groups and centers. We show that these approximation factors match the theoretical lower bounds, assuming standard complexity theory conjectures. Finally, using an (anonymous) open-source implementation, we demonstrate the scalability of our algorithms on large synthetic datasets and assess the price of fairness on real-world data, comparing solution quality with and without fairness constraints.

Large language models (LLMs) are increasingly leveraged as foundational backbones in the development of advanced recommender systems, offering enhanced capabilities through their extensive knowledge and reasoning. Existing LLM-based recommender systems (RSs) often face challenges due to the significant differences between the linguistic semantics of pre-trained LLMs and the collaborative semantics essential for RSs. Typically, these systems apply pre-trained linguistic semantics while learning collaborative semantics from scratch using the LLM-Backbone. However, as LLM architectures are not inherently tailored for recommendation tasks, this approach results in inefficient learning of collaborative information, poor understanding of result correlations, and a failure to leverage traditional RSs features effectively. To address these challenges, we propose LLM-BS, a decoder-only LLM-based generative recommendation framework that integrates endogenous and exogenous Behavioral and Semantic information in a non-intrusive manner. Specifically, we propose 1) a dual-source, knowledge-rich item indexing scheme that integrates indexing sequences for exogenous signals, enabling efficient link-wide processing; 2) a multi-scale reconfiguration alignment that non-intrusively guides the model toward a deeper understanding of both collaborative and semantic signals; 3) an Annealing Adapter designed to finely balance the model’s recommendation performance with its comprehension capabilities. We demonstrate LLM-BS’s effectiveness through rigorous testing on three public benchmarks.

Web applications often include third-party content and scripts to personalize a user's online experience. These scripts have unrestricted access to a user's private data stored in the browser's persistent storage like cookies, localstorage and IndexedDB, associated with the host page. Various mechanisms have been implemented to restrict access to these storage objects, e.g., content security policy, the HttpOnly attribute with cookies, etc. However, the existing mechanisms provide an all-or-none access and do not work in scenarios where web applications need to allow controlled access to cookies and localstorage objects by third-party scripts. If some of these scripts behave maliciously, they can easily access and modify private user information that are stored in the browser objects. The goal of our work is to design a mechanism to enforce fine-grained control of persistent storage objects. We perform an empirical study of persistent storage access by third-party scripts on Tranco's top 10,000 websites and find that 89.84% of all cookie accesses, 90.98% of all localstorage accesses and 72.49% of IndexedDB accesses are done by third-party scripts. Our approach enforces least privilege access for third-party scripts on these objects to ensure their security by attaching labels to the storage objects that specify which domains are allowed to read from and write to these objects. We implement our approach on the Firefox browser and show that it effectively blocks scripts from other domains, which are not allowed access, based on these labels, from accessing the storage objects. We show that our enforcement results in some functionality breakage in websites with the default settings, which can be fixed by correctly labeling the storage objects used by the third-party scripts.

Recommender Systems (RS) aim to generate personalized ranked lists for each user and are also evaluated using ranking metrics. Although personalized ranking is a fundamental aspect of RS, this critical property is often overlooked in the design of model architectures. To address this issue, we propose Rankformer, a ranking-inspired recommendation model. The architecture of Rankformer is inspired by the gradient of the ranking objective, embodying a unique (graph) transformer architecture --- it leverages global information from all users and items to produce more informative representations, and employs specific attention weights to guide the evolution of embeddings towards improved ranking performance. We further develop an acceleration algorithm for Rankformer, reducing its complexity to a linear level with respect to the number of positive instances. Extensive experimental results demonstrate that Rankformer outperforms state-of-the-art methods.

We introduce ColaCare, a framework that enhances Electronic Health Record (EHR) modeling through multi-agent collaboration driven by Large Language Models (LLMs). Our approach seamlessly integrates domain-specific expert models with LLMs to bridge the gap between structured EHR data and text-based reasoning. Inspired by the Multidisciplinary Team (MDT) approach used in clinical settings, ColaCare employs two types of agents: DoctorAgents and a MetaAgent, which collaboratively analyze patient data. Expert models process and generate predictions from numerical EHR data, while LLM agents produce reasoning references and decision-making reports within the MDT-driven collaborative consultation framework. The MetaAgent orchestrates the discussion, facilitating consultations and evidence-based debates among DoctorAgents, simulating diverse expertise in clinical decision-making. We additionally incorporate the Merck Manual of Diagnosis and Therapy (MSD) medical guideline within a retrieval-augmented generation (RAG) module for medical evidence support, addressing the challenge of knowledge currency. Extensive experiments conducted on three EHR datasets demonstrate ColaCare's superior performance in clinical mortality outcome and readmission prediction tasks, underscoring its potential to revolutionize clinical decision support systems and advance personalized precision medicine. The code, complete prompt templates, case studies are publicly available at the anonymous link: https://colacare.netlify.app.

Approximate nearest neighbor (ANN) search is fundamental in various applications such as information retrieval. To enhance efficiency, partition-based methods are proposed to narrow the search space by probing partial partitions, yet they face two common issues. First, in the query phase, a widely adopted strategy in existing studies such as IVF is to probe partitions based on the distance ranks of a query to partition centroids. This inevitably leads to irrelevant partition probing, since data distribution is not considered. Second, in the partition construction phase, all the partition-based methods have the boundary problem that separates a query's $k$NN to multiple partitions and produces a long-tailed $k$NN distribution, degrading the optimal $nprobe$ (i.e., the number of probing partitions) and the search efficiency. To address these problems, we propose LIRA, a LearnIng-based queRy-aware pArtition framework. Specifically, we propose a probing model to learn and directly probe the partitions containing the $k$NN of a query. Probing partitions with the model can reduce probing waste and allow for query-aware probing with query-specific $nprobe$. Moreover, we incorporate the probing model into a learning-based redundancy strategy to mitigate the adverse impact of the long-tailed $k$NN distribution on partition probing. Extensive experiments on real-world vector datasets demonstrate the superiority of LIRA in the trade-off among accuracy, latency, and query fan-out. The results show that LIRA consistently reduces the latency and the query fan-out up to 30\%.

The core of the modern recommender systems lies in learning high-quality embedding representations of users and items to investigate their positional relations in the feature space. Unfortunately, data sparsity caused by difficult-to-access interaction data severely limits the effectiveness of recommender systems. Faced with such a dilemma, various types of self-supervised learning methods have been introduced into recommender systems in an attempt to alleviate the data sparsity through distribution modeling or data augmentation. However, most data augmentation relies on elaborate manual design, which is not only not universal, but the bloated and redundant augmentation process may significantly slow down model training progress. To tackle these limitations, we propose a novel Dual Mixing-based Recommendation Framework (MixRec) to empower data augmentation as we wish. Specifically, we propose individual mixing and collective mixing, respectively. The former aims to provide a new positive sample that is unique to the target (user or item) and to make the pair-wise recommendation loss benefit from it, while the latter aims to portray a new sample that contains group properties in a batch. The two mentioned mixing mechanisms allow for data augmentation with only one parameter that does not need to be set multiple times and can be done in linear time complexity. Besides, we propose the dual-mixing contrastive learning to maximize the utilization of these new-constructed samples to enhance the consistency between pairs of positive samples. Experimental results on four real-world datasets demonstrate the effectiveness of MixRec in terms of recommendation performance, training efficiency, sparsity resistance, and usability.

Information disclosure can compromise privacy when revealed information is correlated with private information. We consider the notion of inferential privacy, which measures privacy leakage by bounding the inferential power a Bayesian adversary can gain by observing a released signal. Our goal is to devise an inferentially-private private information structure that maximizes the informativeness of the released signal, following the Blackwell ordering principle, while adhering to inferential privacy constraints. To achieve this, we devise an efficient release mechanism that achieves the inferentially-private Blackwell optimal private information structure for the setting where the private information is binary. Additionally, we propose a programming approach to compute the optimal structure for general cases given the utility function. The design of our mechanisms builds on our geometric characterization of the Blackwell-optimal disclosure mechanisms under privacy constraints, which may be of independent interest.

The rapid proliferation of fake news on social media threatens social stability, creating an urgent demand for more effective detection methods. While many promising approaches have emerged, most rely on content analysis with limited semantic depth, leading to suboptimal comprehension of news content. To address this limitation, capturing broader-range semantics is essential yet challenging, as it introduces two primary types of noise: fully connecting sentences in news graphs often adds unnecessary structural noise, while highly similar but authenticity-irrelevant sentences introduce feature noise, complicating the detection process. To tackle these issues, we propose BREAK, a broad-range semantics model for fake news detection that leverages a fully connected graph to capture comprehensive semantics while employing dual denoising modules to minimize both structural and feature noise. The semantic structure denoising module balances the graph’s connectivity by iteratively refining it between two bounds: a sequence-based structure as a lower bound and a fully connected graph as the upper bound. This refinement uncovers label-relevant semantic interrelations structures. Meanwhile, the semantic feature denoising module reduces noise from similar semantics by diversifying representations, aligning distinct outputs from the denoised graph and sequence encoders using KL-divergence to achieve feature diversification in high-dimensional space. The two modules are jointly optimized in a bi-level framework, enhancing the integration of denoised semantics into a comprehensive representation for detection. Extensive experiments across four datasets demonstrate that BREAK significantly outperforms existing methods in identifying fake news. Code is available at https://anonymous.4open.science/r/BREAK.