WWW 2025 论文列表

Instant Messaging is a popular mean for daily communication, allowing users to send text and stickers. As the saying goes, "a picture is worth a thousand words", so developing an effective sticker retrieval technique is crucial for enhancing user experience. However, existing sticker retrieval methods rely on labeled data to interpret stickers, and general-purpose Vision-Language Models (VLMs) often struggle to capture the unique semantics of stickers. Additionally, relevant-based sticker retrieval methods lack personalization, creating a gap between diverse user expectations and retrieval results. To address these, we propose the Personalized Sticker Retrieval with Vision-Language Model framework, namely PerSRV, structured into offline calculations and online processing modules. The online retrieval part follows the paradigm of relevant recall and personalized ranking, supported by the offline pre-calculation parts, which are sticker semantic understanding, utility evaluation and personalization modules. Firstly, for sticker-level semantic understanding, we supervised fine-tuned LLaVA-1.5-7B to generate human-like sticker semantics, complemented by textual content extracted from figures and historical interaction queries. Secondly, we investigate three crowd-sourcing metrics for sticker utility evaluation. Thirdly, we cluster style centroids based on users’ historical interactions to achieve personal preference modeling. Finally, we evaluate our proposed PerSRV method on a public sticker retrieval dataset from WeChat, containing 543,098 candidates and 12,568 interactions. Experimental results show that PerSRV significantly outperforms existing methods in multi-modal sticker retrieval. Additionally, our fine-tuned VLM delivers notable improvements in sticker semantic understandings. The code is annoymously available.

Two-tower models are widely adopted in the industrial-scale matching stage across a broad range of application domains, such as content recommendations, advertisement systems, and search engines. This model efficiently handles large-scale candidate item screening by separating user and item representations. However, the decoupling network also leads to a neglect of potential information interaction between the user and item representations. Current state-of-the-art (SOTA) approaches include adding a shallow fully connected layer(i.e., COLD), which is limited by performance and can only be used in the ranking stage. For performance considerations, another approach attempts to capture historical positive interaction information from the other tower by regarding them as the input features(i.e., DAT). Later research showed that the gains achieved by this method are still limited because of lacking the guidance on the next user intent. To address the aforementioned challenges, we propose a "cross-interaction decoupling architecture" within our matching paradigm. This user-tower architecture leverages a diffusion module to reconstruct the next positive intention representation and employs a mixed-attention module to facilitate comprehensive cross-interaction. During the next positive intention generation, we further enhance the accuracy of its reconstruction by explicitly extracting the temporal drift within user behavior sequences. Experiments on two real-world datasets and one industrial dataset demonstrate that our method outperforms the SOTA two-tower models significantly, and our diffusion approach outperforms other generative models in reconstructing item representations. Please find our open-source code repository at the following link: https://anonymous.4open.science/r/T2Diff_ID296/README.md.

Federated learning (FL) has gained widespread attention for its privacy-preserving and collaborative learning capabilities. Due to significant statistical heterogeneity, traditional FL struggles to generalize a shared model across diverse data domains. Personalized federated learning addresses this issue by dividing the model into a globally shared part and a locally private part, with the local model correcting representation biases introduced by the global model. Nevertheless, locally converged parameters more accurately capture domain-specific knowledge, and current methods overlook the potential benefits of these parameters. To address these limitations, we propose PM-MoE architecture. This architecture integrates a mixture of personalized modules and an energy-based personalized modules denoising, enabling each client to select beneficial personalized parameters from other clients. We applied the PM-MoE architecture to nine recent model-split-based personalized federated learning algorithms, achieving performance improvements with minimal additional training. Extensive experiments on six widely adopted datasets and two heterogeneity settings validate the effectiveness of our approach. The source code is available at \url{https://anonymous.4open.science/r/PM-MOE-8315}.

Industrial recommendation systems typically involve a two-stage process: retrieval and ranking, which aims to match users with millions of items. In the retrieval stage, classic embedding-based retrieval (EBR) methods depend on effective negative sampling techniques to enhance both performance and efficiency. However, existing techniques often suffer from false negatives, high cost for sampling quality and semantic information deficiency. To address these limitations, we propose Effective and Semantic-Aware Negative Sampling (ESANS), which integrates two key components: Effective Dense Interpolation Strategy (EDIS) and Multimodal Semantic-Aware Clustering (MSAC). EDIS generates virtual samples within the low-dimensional embedding space to improve the diversity and density of the sampling distribution while minimizing computational costs. MSAC refines the negative sampling distribution by hierarchically clustering item representations based on multimodal information (visual, textual, behavioral), ensuring semantic consistency and reducing false negatives. Extensive offline and online experiments demonstrate the superior efficiency and performance of ESANS.

Heterogeneous Graph Neural Networks (HGNNs) are extensively applied in modeling web-based applications that involve heterogeneous graph structures. Explanation models for HGNNs aim to address their "black box" nature. Enhancing the interpretability of HGNNs leads to a better understanding and can potentially improve predictive performance. However, existing post-hoc HGNN explanation methods cannot impact the HGNN's predictions. Self-explainable homogeneous models also perform poorly on heterogeneous graphs. To address these challenges, we present a Self-Explainable Heterogeneous Graph Neural Network (SEHG), a novel architecture that integrates explanation generation into the learning process of HGNN through two alternative stages. The first stage focuses on producing high-quality explanations while providing predictions alongside. The second stage enhances prediction accuracy by a contrastive learning strategy. Unlike the current methods that rely on manually defined metapaths for structural explanations, SEHG generates important structure and feature explanations by learnable heterogeneous masks. To ensure high-quality and sparsity explanation, these masks are regulated by a uniquely designed range-based penalty during training. Moreover, we introduce HetBA, a collection of synthetic heterogeneous datasets designed to quantify and visualize explanations or heterogeneous graphs. Extensive experiments demonstrate the effectiveness of SEHG, which surpasses strong baselines in real-world node classification tasks by notable margins of up to 3.91%. SEHG also achieves state-of-the-art performance on synthetic datasets with improvement of up to 9.44%, and records the highest fidelity scores in explanation tasks, improving by up to 46.57%. To our knowledge, SEHG is a pioneering self-explainable HGNN framework that achieves state-of-the-art performance on both heterogeneous graph explanation and prediction tasks.

Automating the synthesis of User Interface (UI) plays an important role in enhancing productivity, ensuring design consistency, and expediting the development lifecycle. Recently, the rapid development of Multimodal Large Language Model (MLLM) has made it possible to generate front-end Hypertext Markup Language (HTML) code directly from webpage designs. However, real-world web- pages encompass not only a diverse array of HTML tags but also complex stylesheets, resulting in significantly lengthy code. The lengthy code challenges the performance and efficiency of MLLMs, especially in capturing UI’s structure information. To address this challenge, we propose UICopilot, a structure-aware HTML code generation framework from webpage designs via hierarchy code generation. Our framework introduces a structure model and a code agent, decoupling the generation of the HTML code’s hierarchical structure from its fine-grained details, thereby significantly reducing the model’s burden in producing lengthy code. We evaluate our framework on real-world test datasets, and the experimental results demonstrate that it significantly outperforms existing baselines in both automatic metrics and human evaluations. Specifically, statistical analysis reveals that the majority of human annotators prefer the webpages generated by our framework over those produced by GPT-4V.

Stress haunts people in modern society, which may cause severe health issues if left unattended. With social media becoming an integral part of daily life, leveraging social media to detect stress has gained increasing attention. While the majority of the work focuses on classifying stress states and stress categories, this study introduce a new task aimed at estimating more specific stressors (like exam, writing paper, etc.) through users' posts on social media. Unfortunately, the diversity of stressors with many different classes but a few examples per class, combined with the consistent arising of new stressors over time, hinders the machine understanding of stressors. To this end, we cast the stressor estimation problem within a practical scenario few-shot learning setting, and propose a novel meta-learning based stressor estimation framework that is enhanced by a meta-knowledge inheritance mechanism. This model can not only learn generic stressor context through meta-learning, but also has a good generalization ability to estimate new stressors with little labeled data. A fundamental breakthrough in our approach lies in the inclusion of the meta-knowledge inheritance mechanism, which equips our model with the ability to prevent catastrophic forgetting when adapting to new stressors. The experimental results show that our model achieves state-of-the-art performance compared with the baselines. Additionally, we construct a social media-based stressor estimation dataset that can help train web mining models to facilitate human well-being.

With the development of social media, people are exposed to a vast amount of unverified information, making fact-checking particularly important. Existing fact-checking methods primarily encourage breaking down claims into more easily solvable sub-tasks, and deriving final answers through reasoning with external evidence. However, these models face logical issues regarding whether and how the sub-tasks can logically be combined to form the original claims, and encounter causal errors in the reasoning process due to insufficient evidence or hallucinations from LLMs. In addition, they often suffer from a lack of interpretability. In this paper, we propose $\textbf{Lo}$gical and $\textbf{Ca}$usa$\textbf{l}$ fact-checking (LoCal), a novel fact-checking framework based on multiple LLM-based agents. The usage of multi-agent systems is due to their increasingly demonstrated ability to perform complex tasks in a manner similar to humans. LoCal primarily consists of a decomposing agent, multiple reasoning agents, and two evaluating agents. Specifically, the decomposing agent first utilizes the in-context learning ability of LLMs to break down complex claims into simpler sub-tasks, including fact verification tasks and question answering tasks. Afterwards, two types of reasoning agents are respectively utilized to retrieve external knowledge to address the fact verification tasks that requires comparative analysis skills, and the question answering tasks that necessitates the ability of information extraction from evidence. We then combine the sub-tasks and their corresponding responses to generate a solution for evaluation. In order to enhance logical and causal consistency, two evaluating agents are respectively employed to examine whether the generated solution is logically equivalent to the original claim and determine whether the solution still hold when challenged by the counterfactual label. The evaluating agents provide confidence degrees for the solutions based on the evaluation results and iteratively correct the logical and causal errors in the reasoning process. We evaluate LoCal on two challenging datasets, and the results show that LoCal significantly outperforms all the baseline models across different settings of evidence availability. In addition, LoCal offers better interpretability by providing a structured solution along with detailed evaluating processes. We believe LoCal will provide valuable insights for future agent-based misinformation detection.

Quantitative information about entities constitutes a significant portion of tabular data in open sources and data lakes. Tuch tables often lack consistent labeling and proper schema, posing significant challenges for querying and integration. This paper studies the problem of numerical column annotation in scenarios where quantitative data may be gathered from different sources and unit consistency is a concern. For instance, weight measurements may vary between entities, expressed in kilograms for some and pounds for others, with no accompanying unit information. We investigate the conditions for effectively annotating mixed-unit numeric data, introduce a benchmark for such an annotation task, and propose an algorithm that reliably detects semantic types (e.g., height) and links them to the corresponding types present in a knowledge graph. Our evaluation on a diverse set of columns with mixed units and varying levels of annotation difficulty shows that our method significantly outperforms strong baselines such as GPT-4o-mini and SAND in terms of accuracy, excelling in both detecting mixed units and annotating them with appropriate semantic labels. All our code and data will be publicly released upon acceptance of the paper.

Recent advances in recommender systems have shown that user-system interaction essentially formulates long-term optimization problems, and online reinforcement learning can be adopted to improve recommendation performance. The general solution framework incorporates a value function that estimates the user's expected cumulative rewards in the future and guides the training of the recommendation policy. To avoid local maxima, the policy may explore potential high-quality actions during inference to increase the chance of finding better future rewards. To accommodate the stepwise recommendation process, one widely adopted approach to learning the value function is learning from the difference between the values of two consecutive states of a user. However, we argue that this paradigm involves an incorrect approximation in the stochastic process. Specifically, between the current state and the next state in each training sample, there exist two separate random factors from the stochastic policy and the uncertain user environment. Original TD learning under these mixed random factors may result in a suboptimal estimation of the long-term rewards. As a solution, we show that these two factors can be separately approximated by decomposing the original temporal difference loss. The disentangled learning framework can achieve a more accurate estimation with faster learning and improved robustness against action exploration. As empirical verification of our proposed method, we conduct offline experiments with online simulated environments built based on public datasets.

Graphs are able to model interconnected entities in many online services, supporting a wide range of applications on the Web. This raises an important question: How can we train a graph foundational model on multiple source domains and adapt to an unseen target domain? A major obstacle is that graphs from different domains often exhibit divergent characteristics. Some studies leverage large language models to align multiple domains based on textual descriptions associated with the graphs, limiting their applicability to text-attributed graphs. For text-free graphs, very few recent works attempt to align different feature distributions across domains, while generally neglecting structural differences. In this work, we propose a novel Structure Alignment framework for text-free Multi-domain Graph Pre-Training and cross-domain adaptation (SAMGPT). It is designed to learn multi-domain knowledge from graphs originating in multiple source domains, which can then be adapted to address applications in an unseen target domain. Specifically, we introduce a set of structure tokens to harmonize structure-based aggregation across source domains during the pre-training phase. Next, for cross-domain adaptation, we design dual prompts, namely, holistic prompts and specific prompts, which adapt unified multi-domain structural knowledge and fine-grained, domain-specific information, respectively, to a target domain. Finally, we conduct comprehensive experiments on seven public datasets to evaluate and analyze the effectiveness of SAMGPT. (Codes and data are available at https://anonymous.4open.science/r/SAMGPT for anonymous review.)

Retrieval plays a fundamental role in recommendation systems, search, and natural language processing (NLP) by efficiently finding relevant items from a large corpus given a query. Dot products have been widely used as the similarity function in such retrieval tasks, thanks to Maximum Inner Product Search (MIPS) that enabled efficient retrieval based on dot products. However, state-of-the-art retrieval algorithms have migrated to learned similarities. Such algorithms vary in form; the queries can be represented with multiple embeddings, complex neural networks can be deployed, the item ids can be decoded directly from queries using beam search, and multiple approaches can be combined in hybrid solutions. Unfortunately, we lack efficient solutions for retrieval in these state-of-the-art setups. Our work investigates techniques for efficient retrieval with expressive learned similarity functions. We first prove that Mixture-of-Logits (MoL) is a universal approximator, and can express all learned similarity functions. We then demonstrate how to apply MoL to common retrieval tasks in recommendation systems and NLP. We next propose techniques to retrieve the approximate top $K$ results using MoL with a tight bound. We finally compare our techniques with existing approaches, showing that MoL, with a new mutual information-based load balancing loss we propose, sets new state-of-the-art results across heterogeneous scenarios, including sequential retrieval models in recommendation systems and finetuning language models for question answering; and our approximate top-k retrieval with learned similarities outperforms baselines by up to 105× in latency, while achieving > .99 recall rate of exact algorithms.

WebGL has long been the prevalent API for GPU-accelerated graphics in web browsers, boosting 2D/3D graphical web applications. Despite widespread adoption, WebGL's programming model hinders its rendering performance on modern GPU hardware. To this end, WebGPU has been proposed as the next-generation API of GPU-accelerated processing in web browsers, exhibiting higher performance than WebGL. However, considering the complex logic of WebGL applications and the still-evolving WebGPU specification, statically migrating existing WebGL applications to WebGPU from source code is labor-intensive. To address this issue, we propose GL2GPU, an intermediate layer that dynamically translates WebGL to WebGPU at JavaScript runtime to improve rendering performance. GL2GPU addresses the inconsistencies between the WebGL and WebGPU programming models by emulating WebGL rendering states and leverages performance optimization mechanisms introduced by WebGPU to reduce the overhead of dynamic translation. Evaluation of three representative WebGL benchmarks shows that GL2GPU significantly enhances end-to-end rendering performance while maintaining visual consistency, achieving an average frame time reduction of 45.05\% across different devices and operating systems.

Recommender systems are vital for personalizing user experiences, yet they often rely on implicit feedback data that can be noisy and misleading. Existing denoising studies typically involve either incorporating auxiliary information or learning denoising strategies from interaction data. Nonetheless, they face challenges due to the inherent limitations of external knowledge and interaction data, as well as the non-universality of certain predefined assumptions, which hinder their ability to accurately identify noise. Recently, large language models (LLMs) have garnered significant attention due to their extensive world knowledge and powerful reasoning capabilities. Despite this, the potential of LLMs to enhance the denoising process in recommendations remains largely unexplored. In this paper, we introduce LLaRD, a novel framework that leverages LLMs to improve the denoising process in recommender systems, thereby enhancing overall recommendation performance. Specifically, LLaRD generates denoising-related knowledge by first enriching semantic insights from observational data through LLMs, facilitating a comprehensive inference of user-item preference knowledge. It then employs a novel Chain-of-Thought (CoT) technique over user-item interaction graphs to uncover relation knowledge pertinent to denoising. Finally, it utilizes the Information Bottleneck (IB) principle to align the denoising knowledge generated by LLMs with the recommendation targets, effectively filtering out both data noise and irrelevant knowledge produced by the LLMs. Empirical results demonstrate the effectiveness of our proposed framework, showcasing its superior performance in denoising and recommendation accuracy.

Damage analysis in social media platforms such as Twitter is a comprehensive problem which involves different subtasks for mining damage-related information from tweets e.g., informativeness, humanitarian categories and severity assessment). The comprehensive information obtained by damage analysis enables to identify breaking events around the world in real-time and hence provides aids in emergency responses. Recently, with the rapid development of web technologies, multimodal damage analysis has received increasing attentions due to users' preference of posting multimodal information in social media. Multimodal damage analysis leverages the associated image modality to improve the identification of damage-related information in social media. However, existing works on multimodal damage analysis address each damage-related subtask individually and do not consider their joint training mechanism. In this work, we propose the Bidirectional Multi-task Cascaded multimodal Fusion (BiMCF) approach towards joint multimodal damage analysis. To this end, we introduce the cascaded multimodal fusion framework to separately integrate effective visual and text information for each task, considering that different tasks attend to different information. To exploit the interactions across tasks, bidirectional propagation of the attended image-text interactive information is implemented between tasks, which can lead to enhanced multimodal fusion. Comprehensive experiments are conducted to validate the effectiveness of the proposed approach.

Collaborative Filtering (CF) methods dominate real-world recommender systems given their ability to learn high-quality, sparse ID-embedding tables that effectively capture user preferences. These tables scale linearly with the number of users and items, and are trained to ensure high similarity between embeddings of interacted user-item pairs, while maintaining low similarity for non-interacted pairs. Despite their high performance, encouraging dispersion for non-interacted pairs necessitates expensive regularization (e.g., negative sampling), hurting runtime and scalability. Existing research tends to address these challenges by simplifying the learning process, either by reducing model complexity or sampling data, trading performance for runtime. In this work, we move beyond model-level modifications and study the properties of the embedding tables under different learning strategies. Through theoretical analysis, we find that the singular values of the embedding tables are intrinsically linked to different CF loss functions. These findings are empirically validated on real-world datasets, demonstrating the practical benefits of higher stable rank -- a continuous version of matrix rank which encodes the distribution of singular values. Based on these insights, we propose an efficient warm-start strategy that regularizes the stable rank of the user and item embeddings. We show that stable rank regularization during early training phases can promote higher-quality embeddings, resulting in training speed improvements of up to 65.9%. Additionally, stable rank regularization can act as a proxy for negative sampling, allowing for performance gains of up to 21.2% over loss functions with small negative sampling ratios. Overall, our analysis unifies current CF methods under a new perspective -- their optimization of stable rank -- motivating a flexible regularization method that is easy to implement, yet effective at enhancing CF systems.

Being able to query online public knowledge graphs such as Wikidata or DBpedia is extremely valuable. However, these queries can be interrupted due to the fair use policies enforced by SPARQL endpoint providers, leading to incomplete results. While these policies help maintain responsiveness for public SPARQL endpoints, they compromise the completeness of query results, limiting the feasibility of various downstream tasks. Ideally, we shouldn’t have to choose between completeness and responsiveness. To address this issue, we introduce the concept of SPARQL continuation queries. When a SPARQL endpoint interrupts a query, it returns partial results along with a SPARQL continuation query to retrieve the remaining results. If the continuation query is also interrupted, the process repeats, generating further continuation queries until the complete results are obtained. In our experimention, we show that our continuation server Passage ensures completeness and responsiveness with performances in execution time similar to BlazeGraph.

Existing domain-specific Large Language Models (LLMs) are typically developed by fine-tuning general-purposed LLMs with large-scale domain-specific corpora. However, training on large-scale corpora often fails to effectively organize domain knowledge of LLMs, leading to fragmented understanding. Inspired by how humans connect concepts and organize knowledge through mind maps, we aim to emulate this approach by using ontology with hierarchical conceptual knowledge to reorganize LLM's domain knowledge. From this perspective, we propose an ontology-driven self-training framework called OntoTune, which aims to align LLMs with ontology through in-context learning, enabling the generation of responses guided by the ontology. We leverage in-context learning to identify whether the LLM has acquired the specific concept's ontology knowledge, and select the entries not yet mastered by LLM as the training set to further align the LLM with ontology. Compare to existing domain LLMs based on newly collected large-scale domain-specific corpora, our OntoTune, which relies on the existing, long-term developed ontology and LLM itself, significantly reduces data maintenance costs and offers improved generalization ability. We conduct our study in the medical domain to evaluate the effectiveness of OntoTune, utilizing a standardized medical ontology, SNOMED CT as our ontology source. Experimental results demonstrate that OntoTune achieves state-of-the-art performance in both in-ontology task hypernym discovery and out-of-ontology task medical domain QA. Moreover, compared to the latest direct ontology injection method TaxoLLaMA, our OntoTune better preserves original knowledge of LLM.

As large language models (LLMs) are gaining increasing popularity across a wide range of web applications, it is of great importance to optimize service-level objectives (SLOs) for LLM inference services to enhance user satisfaction and improve the competitiveness of cloud vendors. In this paper, we observe that adjusting the parameters of LLM inference engines can improve service performance, and the optimal parameter configurations of different services are different. Therefore, we propose SCOOT, an automatic performance tuning system to optimize SLOs for each LLM inference service by tuning the parameters of the inference engine. SCOOT jointly exploits single-objective and multiple-objective Bayesian optimization (BO) techniques to handle various optimization objectives via exploration and exploitation. Moreover, SCOOT prunes the search space with known constraints and adopts a random forest to learn hidden constraints during the tuning process to mitigate invalid exploration. To improve the tuning efficiency, SCOOT utilizes the parallel suggestion to accelerate the tuning process. Extensive experiments demonstrate that SCOOT considerably outperforms existing tuning techniques in SLO optimization while greatly improving the tuning efficiency. SCOOT is universally applicable to various LLM inference engines and is easily expandable to new parameters. Currently, SCOOT has already been implemented in the production environment of a leading international technology company.