Filter bubbles have been studied extensively within the context of online content platforms due to their potential to cause undesirable outcomes such as user dissatisfaction or polarization. With the rise of short-video platforms, the filter bubble has been given extra attention because these platforms rely on an unprecedented use of the recommender system to provide relevant content. In our work, we investigate the deep filter bubble, which refers to the user being exposed to narrow content within their broad interests. We accomplish this using one-year interaction data from a top short-video platform in China, which includes hierarchical data with three levels of categories for each video. We formalize our definition of a "deep" filter bubble within this context, and then explore various correlations within the data: first understanding the evolution of the deep filter bubble over time, and later revealing some of the factors that give rise to this phenomenon, such as specific categories, user demographics, and feedback type. We observe that while the overall proportion of users in a filter bubble remains largely constant over time, the depth composition of their filter bubble changes. In addition, we find that some demographic groups that have a higher likelihood of seeing narrower content and implicit feedback signals can lead to less bubble formation. Finally, we propose some ways in which recommender systems can be designed to reduce the risk of a user getting caught in a bubble.
论文检索
输入标题、作者或关键词，从 432 篇学术成果中精准定位
Recommendation systems for Web content distribution intricately connect to the information access and exposure opportunities for vulnerable populations. The emergence of Large Language Models-based Recommendation System (LRS) may introduce additional societal challenges to recommendation systems due to the inherent biases in Large Language Models (LLMs). From the perspective of item-side fairness, there remains a lack of comprehensive investigation into the item-side fairness of LRS given the unique characteristics of LRS compared to conventional recommendation systems. To bridge this gap, this study examines the property of LRS with respect to item-side fairness and reveals the influencing factors of both historical users' interactions and inherent semantic biases of LLMs, shedding light on the need to extend conventional item-side fairness methods for LRS. Towards this goal, we develop a concise and effective framework called IFairLRS to enhance the item-side fairness of an LRS. IFairLRS covers the main stages of building an LRS with specifically adapted strategies to calibrate the recommendations of LRS. We utilize IFairLRS to fine-tune LLaMA, a representative LLM, on MovieLens and Steam datasets, and observe significant item-side fairness improvements. The code can be found in https://github.com/JiangM-C/IFairLRS.git.
In real clinics, the medical data are scattered over multiple hospitals. Due to security and privacy concerns, it is almost impossible to gather all the data together and train a unified model. Therefore, multi-node machine learning systems are currently the mainstream form of model training in healthcare systems. Nevertheless, distributed training relies on the exchange of gradients, which has been proved under the risk of privacy leakage. That means malicious attackers can restore the user's sensitive data by utilizing the publicly shared gradients, which is a serious problem for extremely private data such as Electronic Healthcare Records (EHRs). The performance of the previous gradient attack method will drop rapidly when the batch size of training data increases, which makes it less threatening in practice. However, in this paper, we found in the medical domain, by leveraging prior knowledge like the medical knowledge graph, the leakage risk can be significantly amplified. In particular, we present GraphLeak, which incorporates the medical knowledge graph in gradient leakage attacks. GraphLeak can improve the restoration effect of gradient attacks even under large batches of data. We conduct experimental verification on electronic healthcare record datasets, including eICU and MIMIC-III. Our method has achieved state-of-the-art attack performance compared with previous works. Code is available at https://github.com/anonymous4ai/GraphLeak.
Multimodal recommender systems have acquired applications in broad web scenarios such as e-commerce businesses and short-video platforms. Existing multimodal recommendation methods generally boost performance by introducing item-side multimodal content as supplement information. However, the common training paradigm, i.e., encoding unimodal content respectively and fusing them to fit user preference scores, makes the model biased towards items with prevailing modality content under non-uniform training data. This results in a serious item-side unfairness issue, i.e., some items with prevailing modality content are over-recommended while a large number of items don't receive adequate recommendation opportunities, leaving corresponding content providers at great disadvantage. Aiming to eliminate such modality bias and promote item-side fairness, we propose a fairness-aware modality debiasing framework based on counterfactual inference. In the training stage, we additionally introduce unimodal prediction branches to capture the modality bias. In the inference stage, we conduct a fairness-aware counterfactual inference to adaptively eliminate the modality bias. The proposed framework is model-agnostic and flexible to be implemented in various multimodal recommendation models. Extensive experiments on two datasets demonstrate that the proposed method can significantly enhance item-side fairness while providing competitive recommendation accuracy. Our proposed framework is expected to help mitigate the unfair treatment experienced by vulnerable content providers on multimedia web platforms. Codes are available in https://github.com/tsinghua-fib-lab-WWW2024-Modality-Debiasing.
Food waste and food insecurity are two problems that co-exist worldwide. A major force to combat food waste and insecurity, food rescue platforms (FRP) match food donations to low-resource communities. Since they rely on external volunteers to deliver the food, communicating rescue task difficulty to volunteers is very important for volunteer engagement and retention. We develop a hybrid model with tabular and natural language data to predict the difficulty of a given rescue trip, which significantly outperforms baselines in identifying easy and hard rescues. Furthermore, using storyboards, we conducted interviews with different stakeholders to understand their perspectives on how to integrate such predictions into volunteers' workflow. Motivated by our findings, we developed three explanation methods to generate interpretable insights for volunteers to better understand the predictions. The results from this study are in the process of being adopted at Food Rescue Hero, a large FRP serving over 25 cities across the United States.
Misinformation proliferation on social media platforms is a pervasive threat to the integrity of online public discourse. Genuine users, susceptible to others' influence, often unknowingly engage with, endorse, and re-share questionable pieces of information, collectively amplifying the spread of misinformation. In this study, we introduce an empirical framework to investigate users' susceptibility to influence when exposed to unreliable and reliable information sources. Leveraging two datasets on political and public health discussions on Twitter, we analyze the impact of exposure on the adoption of information sources, examining how the reliability of the source modulates this relationship. Our findings provide evidence that increased exposure augments the likelihood of adoption. Users tend to adopt low-credibility sources with fewer exposures than high-credibility sources, a trend that persists even among non-partisan users. Furthermore, the number of exposures needed for adoption varies based on the source credibility, with extreme ends of the spectrum (very high or low credibility) requiring fewer exposures for adoption. Additionally, we reveal that the adoption of information sources often mirrors users' prior exposure to sources with comparable credibility levels. Our research offers critical insights for mitigating the endorsement of misinformation by vulnerable users, offering a framework to study the dynamics of content exposure and adoption on social media platforms.
Current research concentrates on studying discussions on social media related to structural failures to improve disaster response strategies. However, detecting social web posts discussing concerns about anticipatory failures is under-explored. If such concerns are channeled to the appropriate authorities, it can aid in the prevention and mitigation of potential infrastructural failures. In this paper, we develop an infrastructure ombudsman -- that automatically detects specific infrastructure concerns. Our work considers several recent structural failures in the US. We present a first-of-its-kind dataset of 2,662 social web instances for this novel task mined from Reddit and YouTube.
This paper studies a critical problem of emergent health misinformation detection, aiming to mitigate the spread of misinformation in emergent health domains to support well-informed healthcare decisions towards a Web for good health. Our work is motivated by the lack of timely resources (e.g., medical knowledge, annotated data) during the initial phases of an emergent health event or topic. In this paper, we develop a multi-source domain adaptive framework that jointly exploits medical knowledge and annotated data from different high-resource source domains (e.g., cancer, COVID-19) to detect misleading posts in an emergent target domain (e.g., mpox, polio). Two important challenges exist in developing our solution: 1) how to accurately detect the partially misleading and unverifiable content in an emergent target domain? 2) How to identify the conflicting knowledge facts from different source domains to accurately detect emergent misinformation in the target domain? To address these challenges, we develop MMAdapt, a multi-source multi-class domain adaptive misinformation detection framework that effectively explores diverse knowledge facts from different source domains to accurately detect not only the outright misleading but also the partially misleading or unverifiable posts on the Web. Extensive experimental results on four real-world misinformation datasets demonstrate that MMAdapt substantially outperforms state-of-the-art baselines in accurately detecting misinformation in an emergent health domain.
Online memes have emerged as powerful digital cultural artifacts in the age of social media, offering not only humor but also platforms for political discourse, social critique, and information dissemination. Their extensive reach and influence in shaping online communities' sentiments make them invaluable tools for campaigning and promoting ideologies. Despite the development of several meme generation tools, there remains a gap in their systematic evaluation and their ability to effectively communicate ideologies. Addressing this, we introduce MemeCraft, an innovative meme generator that leverages large language models (LLMs) and visual language models (VLMs) to produce memes advocating specific social movements. MemeCraft presents an end-to-end pipeline, transforming user prompts into compelling multimodal memes without manual intervention. Conscious of the misuse potential in creating divisive content, an intrinsic safety mechanism is embedded to curb hateful meme production. Our assessment, focusing on two UN Sustainable Development Goals-Climate Action and Gender Equality-shows MemeCraft's prowess in creating memes that are both funny and supportive of advocacy goals. This paper highlights how generative AI can promote social good and pioneers the use of LLMs and VLMs in meme generation.
Sketch-based drawing assessments are useful in understanding individuals' cognitive and psychological states, such as cognitive impairment or mental disorders. Hence, these assessments have been developed and applied on a large scale, such as in schools and workplaces, to screen individuals who may require further clinical examination. However, the interpretation of a large number of drawing assessments solely relies on human experts, requiring much time and cost. To address this issue, we introduce a novel scene-level sketch dataset, SceneDAPR, which can be used to automatically analyze the drawing assessment, Draw-A-Person-in-the-Rain (DAPR), a popular psychological drawing assessment used for identifying stressful experiences and coping behavior. The proposed dataset consists of 6,420 objects depicted in 1,399 scene sketches drawn by humans, along with detailed supplementary information about the participants. SceneDAPR includes free-hand drawings from different age groups: children & adolescents, adults, and seniors. Leveraging the proposed SceneDAPR, we develop a web-based drawing assessment system. The extensive experiments demonstrate that our system shows a robust performance across the different age groups in the object detection task as well as a considerable performance compared to human experts. We believe that the proposed new sketch dataset can be used to develop an automatic system for psychological drawing assessments, which can support human experts by reducing the time and cost of analyzing the drawing assessments for a large population. SceneDAPR and experimental code are available at https://github.com/DSAIL-SKKU/SceneDAPR.
In an era dominated by web-based intelligent customer services, the applications of Sentence Pair Matching are profoundly broad. Web agents, for example, automatically respond to customer queries by finding similar past questions, significantly reducing customer service expenses. While current large language models (LLMs) offer powerful text generation capabilities, they often struggle with opacity, potential text toxicity, and difficulty managing domain-specific and confidential business inquiries. Consequently, the widespread adoption of web-based intelligent customer services in real-world business still greatly relies on query-based interactions. In this paper, we introduce a series of model-agnostic techniques aimed at enhancing both the accuracy and interpretability of Chinese pairwise sentence-matching models. Our contributions include (1) An Edit-distance-weighted fine-tuning method, (2) A Bayesian Iterative Prediction algorithm, (3) A Lexical-based Dual Ranking Interpreter, and (4) A Bi-criteria Denoising strategy. Experimental results on the Large-scale Chinese Question Matching Corpus (LCQMC) with a disturbed test demonstrate that our fine-tuning and prediction methods can steadily improve matching accuracy, building on the current state-of-the-art models. Besides, our interpreter with denoising strategy markedly enhances token-level interpretation in rationality and loyalty. In both matching accuracy and interpretation, our approaches outperform classic methods and even LLMs.
Real-world multi-agent systems are often dynamic and continuous, where the agents co-evolve and undergo changes in their trajectories and interactions over time. For example, the COVID-19 transmission in the U.S. can be viewed as a multi-agent system, where states act as agents and daily population movements between them are interactions. Estimating the counterfactual outcomes in such systems enables accurate future predictions and effective decision-making, such as formulating COVID-19 policies. However, existing methods fail to model the continuous dynamic effects of treatments on the outcome, especially when multiple treatments (e.g., "stay-at-home" and "get-vaccine" policies) are applied simultaneously. To tackle this challenge, we propose Causal Graph Ordinary Differential Equations (CAG-ODE), a novel model that captures the continuous interaction among agents using a Graph Neural Network (GNN) as the ODE function. The key innovation of our model is to learn time-dependent representations of treatments and incorporate them into the ODE function, enabling precise predictions of potential outcomes. To mitigate confounding bias, we further propose two domain adversarial learning-based objectives, which enable our model to learn balanced continuous representations that are not affected by treatments or interference. Experiments on two datasets (i.e., COVID-19 and tumor growth) demonstrate the superior performance of our proposed model.
In the face of rising surface temperatures from climate change, impacting biodiversity, extreme weather events, and agricultural productivity, understanding the drivers behind temperature changes is imperative. Traditional global climate models (GCMs) are computationally expensive, limiting their applicability, while machine learning approaches, though promising, face interpretability challenges due to their "black box" nature, especially in a dynamic setting where the data is constantly evolving. We propose DUO, a framework to identify shifts in important features and feature combinations as the data distribution changes over time. Our model independently assesses the importance of features and their interactions while also evaluating their relevance when combined with additional features, contributing to the target class. As a case study, we apply DUO to assess the shifts in climate drivers for station-level temperatures in six locations across New Zealand from 1980 to 2020, we identify specific humidity, geopotential height, and air temperature at high atmospheric pressure levels as the most important features for describing temperature variability. By revealing how climate drivers change over time, DUO contributes to a deeper understanding of temperature change patterns, enabling practitioners to develop targeted and adaptive mitigation strategies.
CapAlign: Improving Cross Modal Alignment via Informative Captioning for Harmful Meme Detection
PDF ↗Harmful memes detection is challenging due to the semantic gap between different modalities. Previous studies mainly focus on feature extraction and fusion to learn discriminative information from memes. However, they ignore the misalignment of the modalities caused by the modality gap and suffer from data scarcity, resulting in insufficient learning of fusion-based models. Recently, researchers transformed images into textual captions and used language models for predictions, resulting in non-informative image captions. To address these gaps, this paper proposes an instructions-based abstracting approach CapAlign, in zero-shot visual question-answering settings. Precisely, we prompt a large language model (LLM) to ask informative questions to a pre-trained vision-language model and use the dialogues to generate a high-quality image caption. Further, to align the generated caption with the textual content of a meme, we used an LLM with instructions to generate informative captions of the meme and then prepend it with the attributes of the visual content of a meme to a prompt-based LM for prediction. Experimental findings on two benchmark datasets show that our approach produces informative captions and outperforms state-of-the-art methods for detecting harmful memes.
In this paper, we address the challenge of detecting hateful memes in the low-resource setting where only a few labeled examples are available. Our approach leverages the compositionality of Low-rank adaptation (LoRA), a widely used parameter-efficient tuning technique. We commence by fine-tuning large language models (LLMs) with LoRA on selected tasks pertinent to hateful meme detection, thereby generating a suite of LoRA modules. These modules are capable of essential reasoning skills for hateful meme detection. We then use the few available annotated samples to train a module composer, which assigns weights to the LoRA modules based on their relevance. The model's learnable parameters are directly proportional to the number of LoRA modules. This modularized network, underpinned by LLMs and augmented with LoRA modules, exhibits enhanced generalization in the context of hateful meme detection. Our evaluation spans three datasets designed for hateful meme detection in a few-shot learning context. The proposed method demonstrates superior performance to traditional in-context learning, which is also more computationally intensive during inference.
The recommender systems on online platforms assist users in finding personalized information, yet this also leads to the issue of limited diversity, potentially giving rise to societal issues such as filter bubbles. Despite significant progress in diversified recommendation algorithms, they have not been extensively experimented with and evaluated for effectiveness in large-scale, full-stage industrial recommender systems. Specifically, industrial recommenders usually consist of three stages of matching, ranking, and re-ranking, in which specific characteristics lead to critical challenges for promoting both recommendation diversity and user engagement. First, user interests are partially observed due to only relevance maximization. Second, item-side feature-aware bias causes imbalanced recommendations. Last, the impact of diversity perception on user engagement stresses the necessity of explicit diversity modeling. To address these challenges in industrial systems, in this work, we deploy several existing diversified algorithms in a real-world short-video platform, including exploration-exploitation, feature-aware debiasing, and diversity optimization. We conduct large-scale online A/B testing for evaluation via online metrics of user engagement and recommendation diversity. Performance improvement across full stages demonstrates the effectiveness of these simple solutions. From comparing performance across different stages and algorithms, we identify that the ranking stage is the most suitable for real-world deployment, and the combination of debiasing and diversity optimization is a promising direction in terms of diversified recommendations. This work provides experiential guidance for the large-scale deployment of diversified algorithms and the construction of a more inclusive platform on the Web.
Multimodal tasks require learning a joint representation of the constituent modalities of data. Contrastive learning learns a joint representation by using a contrastive loss. For example, CLIP takes as input image-caption pairs and is trained to maximize the similarity between an image and its corresponding caption in actual image-caption pairs, while minimizing the similarity for arbitrary image-caption pairs. This approach operates on the premise that the caption depicts the image's content. However, this assumption does not always hold true for tweets that contain both text and images. Previous studies have indicated that the connection between the image and the text in a tweet is more intricate and complex. We study the effectiveness of pre-trained multimodal contrastive learning models, specifically, CLIP, and ALIGN, on the task of classifying multimodal crisis related tweets. Our experiments using two publicly available datasets, CrisisMMD and DMD, show that despite the intricate relationships in tweets, pre-trained contrastive learning models fine-tuned with task-specific data produce better results than prior approaches used for the multimodal classification of crisis related tweets. Additionally, the experiments show that the contrastive learning models are effective in low-data few-shot and cross-domain settings.
Children are among the most vulnerable online populations. Reports of child sexual exploitation on social media and apps have grown annually at an alarming rate and are overwhelming investigators. Even a single case can require examining millions of messages involving hundreds of victims. Triage and prioritization based on victims' experiences is an unfortunate necessity. Using a chat dataset of more than 3 million messages between victims and perpetrators, we evaluate and contribute tools for analyzing the experiences of victims of sexual exploitation. We develop both supervised and unsupervised methods to classify messages into categories of interest to law enforcement, such as age requests, persuasion, and sexual messages. We also introduce a conversation clustering technique to illuminate differences among victims' experiences based on their chat history. Through a qualitative analysis, we demonstrate that the learned clusters are coherent and represent distinct conversation patterns. For example, we can distinguish groups of users who never comply with sexual requests, comply after a few conversations, or comply immediately after being targeted. We expect this approach and associated visualizations will aid law enforcement, industry moderators, and sociologists who need to analyze massive corpora in this domain. Finally, we validate prior models derived from conversations involving adults pretending to be minors and provide statistics that could help undercover adults more accurately portray minor victims.
Human vs ChatGPT: Effect of Data Annotation in Interpretable Crisis-Related Microblog Classification
PDF ↗Recent studies have exploited the vital role of microblogging platforms, such as Twitter, in crisis situations. Various machine-learning approaches have been proposed to identify and prioritize crucial information from different humanitarian categories for preparation and rescue purposes. In crisis domain, the explanation of models' output decisions is gaining significant research momentum. Some previous works focused on human annotations of rationales to train and extract supporting evidence for model interpretability. However, such annotations are usually expensive, require much effort, and are not always available in real-time situations of a new crisis event. In this paper, we investigate the recent advances in large language models (LLMs) as data annotators on informal tweet text. We perform a detailed qualitative and quantitative evaluation of ChatGPT rationale annotations over a few-shot setup. ChatGPT annotations are quite close to humans but less precise in nature. Further, we propose an active learning-based interpretable classification model from a small set of annotated data. Our experiments show that (a). ChatGPT has the potential to extract rationales for the crisis tweet classification tasks, but the performance is slightly less than the model trained on human-annotated rationale data (\sim3-6%), (b). active learning setup can help reduce the burden of manual annotations and maintain a trade-off between performance and data size.
Web resources in linked open data (LOD) are comprehensible to humans through literal textual values attached to them, such as labels, notes, or comments. Word choices in literals may not always be neutral. When culturally stereotyping terminology is used in literals, they may appear as offensive to users in interfaces and propagate stereotypes to algorithms trained on them. We study how frequently and in which literals contentious terms about people and cultures occur in LOD and whether there are attempts to mark the usage of such terms. For our analysis, we reuse English and Dutch terms from a knowledge graph that provides opinions of experts from the cultural heritage domain about terms' contentiousness. We inspect occurrences of these terms in four widely used datasets: Wikidata, The Getty Art & Architecture Thesaurus, Princeton WordNet, and Open Dutch WordNet. Some terms are ambiguous and contentious only in particular senses. Applying word sense disambiguation, we generate a set of literals relevant to our analysis. We found that contentious terms frequently appear in descriptive and labelling literals, such as preferred labels that are usually displayed in interfaces and used for indexing. In some cases, LOD contributors mark contentious terms with words and phrases in literals (implicit markers) or properties linked to resources (explicit markers). However, such marking is rare and non-consistent in all datasets. Our quantitative and qualitative insights could be helpful in developing more systematic approaches to address the propagation of stereotypes via LOD.