Generative foundation models (GenFMs), including large language and multimodal models, are transforming information retrieval and knowledge management. However, their rapid adoption raises urgent concerns about social responsibility, trustworthiness, and governance. This tutorial offers a comprehensive, hands-on overview of recent advances in responsible GenFMs, covering foundational concepts, multi-dimensional risk taxonomies (including safety, privacy, robustness, truthfulness, fairness, and machine ethics), state-of-the-art evaluation benchmarks, and effective mitigation strategies. We integrate real-world case studies and practical exercises using open-source tools, and present key perspectives from both policy and industry, including recent regulatory developments and enterprise practices. The session concludes with a discussion of open challenges, providing actionable guidance for the CIKM community.

The proposed ''SciSoc LLM Workshop: Large Language Models for Scientific and Societal Advances'' aims to explore the profound implications and potential of Large Language Models (LLMs) in driving forward scientific inquiry and addressing critical societal challenges. As LLMs such as GPT-4 continue to redefine boundaries in both complexity and capability, their integration into the scientific and societal domains is not just beneficial but essential. In particular, LLMs have demonstrated substantial value in improving our understanding of complex datasets and generating insights across various fields such as healthcare, environmental science, education, and public policy. By bringing together experts and enthusiasts from diverse fields, the workshop aims to foster a comprehensive understanding of how LLMs can redefine traditional research methodologies. Participants will explore innovative ways to harness the power of LLMs for greater efficiency and innovation in their respective fields, potentially catalyzing a new era of scientific and societal advancement.

The integration of Large Language Models (LLMs) into the healthcare domain has the potential to significantly enhance patient care and support through the development of empathetic, patient-facing chatbots. This study investigates an intriguing question Can ChatGPT respond with a greater degree of empathy than those typically offered by physicians? To answer this question, we collect a de-identified dataset of patient messages and physician responses from a hospital and generate alternative replies using ChatGPT. We then introduce a set of empathy ranking evaluation (EMRank) metrics to automatically judge the empathy degree. We further conduct human study to gauge the empathy level of responses. Our findings indicate that LLM-powered chatbots have the potential to surpass human physicians in delivering empathetic communication, suggesting a promising avenue for enhancing patient care and reducing professional burnout. To summary, this study not only highlights the importance of clinical empathy in patient interactions but also proposes a set of automatic empathy ranking metrics, paving the way for the broader adoption of LLMs in healthcare.

News media has been frequently utilized as a political tool to stray from facts, making biased statements and claims without evidence. During the COVID-19 vaccine campaign, politically biased news (PBN) has significantly undermined public trust in vaccines. Despite medical evidence showing the benefits of these vaccines, the misperceptions of the vaccine's safety, risks, and efficacy have led to a non-negligible fraction of the population resistant to receiving the vaccine. In this paper, we analyze: (i) how inherent vaccine stances subtly influence individuals' selection of news sources and participation in social media discussions; and (ii) the impact of exposure to PBN on users' attitudes toward vaccines. In doing so, we first curate a comprehensive dataset that connects PBN with related social media discourse. Utilizing advanced deep learning and causal inference techniques, we reveal distinct user behaviors between social media groups with various vaccine stances. Moreover, we observe that individuals with moderate stances, particularly the vaccine-hesitant majority, are more vulnerable to the influence of PBN compared to those with extreme views. Our findings provide critical insights to foster this line of research.

Graph Neural Networks (GNNs) excel in diverse tasks, yet their applications in high-stakes domains are often hampered by unreliable predictions. Although numerous uncertainty quantification methods have been proposed to address this limitation, they often lackrigorous uncertainty estimates. This work makes the first attempt to introduce a distribution-free and model-agnostic uncertainty quantification approach to construct a predictive interval with a statistical guarantee for GNN-based link prediction. We term it asconformalized link prediction. Our approach builds upon conformal prediction (CP), a framework that promises to construct statistically robust prediction sets or intervals. There are two primary challenges: first, given dependent data like graphs, it is unclear whether the critical assumption in CP --- exchangeability --- still holds when applied to link prediction. Second, even if the exchangeability assumption is valid for conformalized link prediction, we need to ensure high efficiency, i.e., the resulting prediction set or the interval length is small enough to provide useful information. To tackle these challenges, we first theoretically and empirically establish a permutation invariance condition for the application of CP in link prediction tasks, along with an exact test-time coverage. Leveraging the important structural information in graphs, we then identify a novel and crucial connection between a graph's adherence to the power law distribution and the efficiency of CP. This insight leads to the development of a simple yet effective sampling-based method to align the graph structure with a power law distribution prior to the standard CP procedure. Extensive experiments demonstrate that for conformalized link prediction, our approach achieves the desired marginal coverage while significantly improving the efficiency of CP compared to baseline methods.

In the rapidly evolving landscape of artificial intelligence, multimodal learning systems (MMLS) have gained traction for their ability to process and integrate information from diverse modality inputs. Their expanding use in vital sectors such as healthcare has made safety assurance a critical concern. However, the absence of systematic research into their safety is a significant barrier to progress in this field. To bridge the gap, we present the first taxonomy that systematically categorizes and assesses MMLS safety. This taxonomy is structured around four fundamental pillars that are critical to ensuring the safety of MMLS: robustness, alignment, monitoring, and controllability. Leveraging this taxonomy, we review existing methodologies, benchmarks, and the current state of research, while also pinpointing the principal limitations and gaps in knowledge. Finally, we discuss unique challenges in MMLS safety. In illuminating these challenges, we aim to pave the way for future research, proposing potential directions that could lead to significant advancements in the safety protocols of MMLS.

We propose a simple yet effective solution to tackle the often-competing goals of fairness and utility in classification tasks. While fairness ensures that the model's predictions are unbiased and do not discriminate against any particular group or individual, utility focuses on maximizing the model's predictive performance. This work introduces the idea of leveraging aleatoric uncertainty (e.g., data ambiguity) to improve the fairness-utility trade-off. Our central hypothesis is that aleatoric uncertainty is a key factor for algorithmic fairness and samples with low aleatoric uncertainty are modeled more accurately and fairly than those with high aleatoric uncertainty. We then propose a principled model to improve fairness when aleatoric uncertainty is high and improve utility elsewhere. Our approach first intervenes in the data distribution to better decouple aleatoric uncertainty and epistemic uncertainty. It then introduces a fairness-utility bi-objective loss defined based on the estimated aleatoric uncertainty. Our approach is theoretically guaranteed to improve the fairness-utility trade-off. Experimental results on both tabular and image datasets show that the proposed approach outperforms state-of-the-art methods w.r.t. the fairness-utility trade-off and w.r.t. both group and individual fairness metrics. This work presents a fresh perspective on the trade-off between utility and algorithmic fairness and opens a key avenue for the potential of using prediction uncertainty in fair machine learning.

Graph Neural Networks (GNNs) are powerful tools for learning representations on graphs, such as social networks. However, their vulnerability to privacy inference attacks restricts their practicality, especially in high-stake domains. To address this issue, privacy-preserving GNNs have been proposed, focusing on preserving node and/or link privacy. This work takes a step back and investigates how GNNs contribute to privacy leakage. Through theoretical analysis and simulations, we identify message passing under structural bias as the core component that allows GNNs to propagate andamplify privacy leakage. Building upon these findings, we propose a principled privacy-preserving GNN framework that effectively safeguards both node and link privacy, referred to as dual-privacy preservation. The framework comprises three major modules: a Sensitive Information Obfuscation Module that removes sensitive information from node embeddings, a Dynamic Structure Debiasing Module that dynamically corrects the structural bias, and an Adversarial Learning Module that optimizes the privacy-utility trade-off. Experimental results on four benchmark datasets validate the effectiveness of the proposed model in protecting both node and link privacy while preserving high utility for downstream tasks, such as node classification.

Online user engagement is highly influenced by various machine learning models, such as recommender systems. These systems recommend new items to the user based on the user’s historical interactions. Implicit recommender systems reflect a binary setting showing whether a user interacted (e.g., clicked on) with an item or not. However, the observed clicks may be due to various causes such as user’s interest, item’s popularity, and social influence factors. Traditional recommender systems consider these causes under a unified representation, which may lead to the emergence and amplification of various biases in recommendations. However, recent work indicates that by disentangling the unified representations, one can mitigate bias (e.g., popularity bias) in recommender systems and help improve recommendation performance. Yet, prior work in causal disentanglement in recommendations does not consider a crucial factor, that is, social influence. Social theories such as homophily and social influence provide evidence that a user’s decision can be highly influenced by the user’s social relations. Thus, accounting for the social relations while disentangling leads to less biased recommendations. To this end, we identify three separate causes behind an effect (e.g., clicks): (a) user’s interest, (b) item’s popularity, and (c) user’s social influence. Our approach seeks to causally disentangle the user and item latent features to mitigate popularity bias in implicit feedback–based social recommender systems. To achieve this goal, we draw from causal inference theories and social network theories and propose a causality-aware disentanglement method that leverages both the user–item interaction network and auxiliary social network information. Experiments on real-world datasets against various state-of-the-art baselines validate the effectiveness of the proposed model for mitigating popularity bias and generating de-biased recommendations.

To meet the growing bike-sharing demands and make people’s travel convenient, the companies need to add new stations at locations where demands exceed supply. Before making reliable decisions on adding new stations, it is required to understand the spatial effects of new stations on the station network. In this paper, we study the deployment of the new station by estimating its varied causal effects on the demands of nearby stations, e.g., how does adding a new station (treatment) causally influence the demands (outcome) of nearby stations? When working with observational data, we should control hidden confounders, which cause spurious relations between treatments and outcomes. However, previous studies use historical data of the individual unit (e.g., the station’s historical demands) to approximate its hidden confounders, which cannot deal with the lack of historical data for new stations. And the conventional methods overlook the differences between units, which cannot be applied to our problem. To overcome the challenges, we propose a novel model (CausalSE) to estimate the varied effects of new stations on nearby stations, which uses the shared knowledge (i.e., similar traveling patterns among stations) to approximate hidden confounders. Experimental results on real-world datasets show that CausalSE outperforms 6 state-of-the-art methods.

Nowadays, fake news easily propagates through online social networks and becomes a grand threat to individuals and society. Assessing the authenticity of news is challenging due to its elaborately fabricated contents, making it difficult to obtain large-scale annotations for fake news data. Due to such data scarcity issues, detecting fake news tends to fail and overfit in the supervised setting. Recently, graph neural networks (GNNs) have been adopted to leverage the richer relational information among both labeled and unlabeled instances. Despite their promising results, they are inherently focused on pairwise relations between news, which can limit the expressive power for capturing fake news that spreads in a group-level. For example, detecting fake news can be more effective when we better understand relations between news pieces shared among susceptible users. To address those issues, we propose to leverage a hypergraph to represent group-wise interaction among news, while focusing on important news relations with its dual-level attention mechanism. Experiments based on two benchmark datasets show that our approach yields remarkable performance and maintains the high performance even with a small subset of labeled news data.

Machine learning algorithms typically assume that the training and test samples come from the same distributions, i.e., in-distribution. However, in open-world scenarios, streaming big data can be Out-Of-Distribution (OOD), rendering these algorithms ineffective. Prior solutions to the OOD challenge seek to identify invariant features across different training domains. The underlying assumption is that these invariant features should also work reasonably well in the unlabeled target domain. By contrast, this work is interested in the domain-specific features that include both invariant features and features unique to the target domain. We propose a simple yet effective approach that relies on correlations in general regardless of whether the features are invariant or not. Our approach uses the most confidently predicted samples identified by an OOD base model (teacher model) to train a new model (student model) that effectively adapts to the target domain. Empirical evaluations on benchmark datasets show that the performance is improved over the SOTA by ∼10-20%.

Increased social media use has contributed to the greater prevalence of abusive, rude, and offensive textual comments. Machine learning models have been developed to detect toxic comments online, yet these models tend to show biases against users with marginalized or minority identities (e.g., females and African Americans). Established research in debiasing toxicity classifiers often (1) takes a static or batch approach, assuming that all information is available and then making a one-time decision; and (2) uses a generic strategy to mitigate different biases (e.g., gender and racial biases) that assumes the biases are independent of one another. However, in real scenarios, the input typically arrives as a sequence of comments/words over time instead of all at once. Thus, decisions based on partial information must be made while additional input is arriving. Moreover, social bias is complex by nature. Each type of bias is defined within its unique context, which, consistent with intersectionality theory within the social sciences, might be correlated with the contexts of other forms of bias. In this work, we consider debiasing toxicity detection as a sequential decision-making process where different biases can be interdependent. In particular, we study debiasing toxicity detection with two aims: (1) to examine whether different biases tend to correlate with each other; and (2) to investigate how to jointly mitigate these correlated biases in an interactive manner to minimize the total amount of bias. At the core of our approach is a framework built upon theories of sequential Markov Decision Processes that seeks to maximize the prediction accuracy and minimize the bias measures tailored to individual biases. Evaluations on two benchmark datasets empirically validate the hypothesis that biases tend to be correlated and corroborate the effectiveness of the proposed sequential debiasing strategy.

Online reviews enable consumers to engage with companies and provide important feedback. Due to the complexity of the high-dimensional text, these reviews are often simplified as a single numerical score, e.g., ratings or sentiment scores. This work empirically examines the causal effects of user-generated online reviews on a granular level: we consider multiple aspects, e.g., the Food and Service of a restaurant. Understanding consumers' opinions toward different aspects can help evaluate business performance in detail and strategize business operations effectively. Specifically, we aim to answer interventional questions such as What will the restaurant popularity be if the quality w.r.t. its aspect Service is increased by 10%? The defining challenge of causal inference with observational data is the presence of "confounder'', which might not be observed or measured, e.g., consumers' preference to food type, rendering the estimated effects biased and high-variance. To address this challenge, we have recourse to the multi-modal proxies such as the consumer profile information and interactions between consumers and businesses. We show how to effectively leverage the rich information to identify and estimate causal effects of multiple aspects embedded in online reviews. Empirical evaluations on synthetic and real-world data corroborate the efficacy and shed light on the actionable insight of the proposed approach.

An important problem in causal inference is to break down the total effect of a treatment on an outcome into different causal pathways and to quantify the causal effect in each pathway. For instance, in causal fairness, the total effect of being a male employee (i.e., treatment) constitutes its direct effect on annual income (i.e., outcome) and the indirect effect via the employee's occupation (i.e., mediator). Causal mediation analysis (CMA) is a formal statistical framework commonly used to reveal such underlying causal mechanisms. One major challenge of CMA in observational studies is handling confounders, variables that cause spurious causal relationships among treatment, mediator, and outcome. Conventional methods assume sequential ignorability that implies all confounders can be measured, which is often unverifiable in practice. This work aims to circumvent the stringent sequential ignorability assumptions and consider hidden confounders. Drawing upon proxy strategies and recent advances in deep learning, we propose to simultaneously uncover the latent variables that characterize hidden confounders and estimate the causal effects. Empirical evaluations using both synthetic and semi-synthetic datasets validate the effectiveness of the proposed method. We further show the potentials of our approach for causal fairness analysis.

Recent studies have shown Graph Neural Networks (GNNs) are extremely vulnerable to attribute inference attacks. To tackle this challenge, existing privacy-preserving GNNs research assumes that the sensitive attributes of all users are known beforehand. However, due to different privacy preferences, some users (i.e., private users) may prefer not to reveal sensitive information that others (i.e., non-private users) would not mind disclosing. For example, in social networks, male users are typically less sensitive to their age information than female users. The age disclosure of male users can lead to the age information of female users in the network exposed. This is partly because social media users are connected, the homophily property and message-passing mechanism of GNNs can exacerbate individual privacy leakage. In this work, we study a novel and practical problem of learning privacy-preserving GNNs with partially observed sensitive attributes.

Recent years have witnessed remarkable progress towards computational fake news detection. To mitigate its negative impact, we argue that it is critical to understand what user attributes potentially cause users to share fake news. The key to this causal-inference problem is to identify confounders -- variables that cause spurious associations between treatments (e.g., user attributes) and outcome (e.g., user susceptibility). In fake news dissemination, confounders can be characterized by fake news sharing behavior that inherently relates to user attributes and online activities. Learning such user behavior is typically subject to selection bias in users who are susceptible to share news on social media. Drawing on causal inference theories, we first propose a principled approach to alleviating selection bias in fake news dissemination. We then consider the learned unbiased fake news sharing behavior as the surrogate confounder that can fully capture the causal links between user attributes and user susceptibility. We theoretically and empirically characterize the effectiveness of the proposed approach and find that it could be useful in protecting society from the perils of fake news.

There are many scenarios where short- and long-term causal effects of an intervention are different. For example, low-quality ads may increase short-term ad clicks but decrease the long-term revenue via reduced clicks. This work, therefore, studies the the problem of long-term effect where the outcome of primary interest, orprimary outcome, takes months or even years to accumulate. The observational study of long-term effect presents unique challenges. First, the confounding bias causes large estimation error and variance, which can further accumulate towards the prediction of primary outcomes. Second, short-term outcomes are often directly used as the proxy of the primary outcome, i.e., thesurrogate. Nevertheless, this method entails the strong surrogacy assumption that is often impractical. To tackle these challenges, we propose to build connections between long-term causal inference and sequential models in machine learning. This enables us to learnsurrogate representations that account for thetemporal unconfoundedness and circumvent the stringent surrogacy assumption by conditioning on the inferred time-varying confounders. Experimental results show that the proposed framework outperforms the state-of-the-art.

Cyberbullying, identified as intended and repeated online bullying behavior, has become increasingly prevalent in the past few decades. Despite the significant progress made thus far, the focus of most existing work on cyberbullying detection lies in the independent content analysis of different comments within a social media session. We argue that such leading notions of analysis suffer from three key limitations: they overlook the temporal correlations among different comments; they only consider the content within a single comment rather than the topic coherence across comments; they remain generic and exploit limited interactions between social media users. In this work, we observe that user comments in the same session may be inherently related, e.g., discussing similar topics, and their interaction may evolve over time. We also show that modeling such topic coherence and temporal interaction are critical to capture the repetitive characteristics of bullying behavior, thus leading to better predicting performance. To achieve the goal, we first construct a unified temporal graph for each social media session. Drawing on recent advances in graph neural network, we then propose a principled graph-based approach for modeling the temporal dynamics and topic coherence throughout user interactions. We empirically evaluate the effectiveness of our approach with the tasks of session-level bullying detection and comment-level case study. Our code is released to public. 1

Social media is a vital means for information-sharing due to its easy access, low cost, and fast dissemination characteristics. However, increases in social media usage have corresponded with a rise in the prevalence of cyberbullying. Most existing cyberbullying detection methods aresupervised and, thus, have two key drawbacks: (1) The data labeling process is often time-consuming and labor-intensive; (2) Current labeling guidelines may not be generalized to future instances because of different language usage and evolving social networks. To address these limitations, this work introduces a principled approach forunsupervised cyberbullying detection. The proposed model consists of two main components: (1) Arepresentation learning network that encodes the social media session by exploiting multi-modal features, e.g., text, network, and time. (2) Amulti-task learning network that simultaneously fits the comment inter-arrival times and estimates the bullying likelihood based on a Gaussian Mixture Model. The proposed model jointly optimizes the parameters of both components to overcome the shortcomings of decoupled training. Our core contribution is an unsupervised cyberbullying detection model that not only experimentally outperforms the state-of-the-art unsupervised models, but also achieves competitive performance compared to supervised models.

Deep architectures are trained on massive amounts of labeled data to guarantee the performance of classification. In the absence of labeled data, domain adaptation often provides an attractive option given that labeled data of a similar nature but from a different domain is available. Previous work has chiefly focused on learning domain invariant representations but overlooked the issues of label imbalance in a single domain or across domains, which are common in many machine learning applications such as fake news detection. In this paper, we study a new cross-domain classification problem where data in each domain can be imbalanced (data imbalance), i.e., the classes are not evenly distributed, and the ratio of the number of positive over negative samples varies across domains (domain imbalance). This cross-domain problem is challenging as it entails covariate bias in the input feature space and representation bias in the latent space where domain invariant representations are learned. To address the challenge, in this paper, we propose an effective approach that leverages a doubly balancing strategy to simultaneously control these two types of bias and learn domain invariant representations. To this end, the proposed method aims to learn representations that are (i) robust to data and domain imbalance, (ii) discriminative between classes, and (iii) invariant across domains. Extensive evaluations of two important real-world applications corroborate the effectiveness of the proposed framework.

Cyberbullying has become one of the most pressing online risks for young people and has raised serious concerns in society. The emerging literature identifies cyberbullying as repetitive acts that occur over time rather than one-off incidents. Yet, there has been relatively little work to model the hierarchical structure of social media sessions and the temporal dynamics of cyberbullying in online social network sessions. We propose a hierarchical attention network for cyberbullying detection that takes these aspects of cyberbullying into account. The primary distinctive characteristics of our approach include: (i) a hierarchical structure that mirrors the structure of a social media session; (ii) levels of attention mechanisms applied at the word and comment level, thereby enabling the model to pay different amounts of attention to words and comments, depending on the context; and (iii) a cyberbullying detection task that also predicts the interval of time between two adjacent comments. These characteristics allow the model to exploit the commonalities and differences across these two tasks to improve the performance of cyberbullying detection. Experiments on a real-world dataset from Instagram, the social media platform on which the highest percentage of users have reported experiencing cyberbullying, reveal that the proposed architecture outperforms the state-of-the-art method.

Over the last decade, research has revealed the high prevalence of cyberbullying among youth and raised serious concerns in society. Information on the social media platforms where cyberbullying is most prevalent (e.g., Instagram, Facebook, Twitter) is inherently multi-modal, yet most existing work on cyberbullying identification has focused solely on building generic classification models that rely exclusively on text analysis of online social media sessions (e.g., posts). Despite their empirical success, these efforts ignore the multi-modal information manifested in social media data (e.g., image, video, user profile, time, and location), and thus fail to offer a comprehensive understanding of cyberbullying. Conventionally, when information from different modalities is presented together, it often reveals complementary insights about the application domain and facilitates better learning performance. In this paper, we study the novel problem of cyberbullying detection within a multi-modal context by exploiting social media data in a collaborative way. This task, however, is challenging due to the complex combination of both cross-modal correlations among various modalities and structural dependencies between different social media sessions, and the diverse attribute information of different modalities. To address these challenges, we propose XBully, a novel cyberbullying detection framework, that first reformulates multi-modal social media data as a heterogeneous network and then aims to learn node embedding representations upon it. Extensive experimental evaluations on real-world multi-modal social media datasets show that the XBully framework is superior to the state-of-the-art cyberbullying detection models.