Forming effective expert teams is central to domains where solving complex problems requires diverse, complementary skills. However, automating this task is highly challenging due to sparse co-occurrence data, long-tailed expert participation, and the combinatorial complexity of unseen skill configurations. Existing graph-based, probabilistic, and neural approaches often struggle with generalization, fairness, and robustness, leading to biased selections that favor historically popular experts over more suitable candidates. To address these challenges, we propose a generative framework for expert team formation based on denoising diffusion probabilistic models. We cast team formation as skill-conditioned imputation (i.e., inpainting), where skills are treated as observed context and the expert component is generated via conditional diffusion sampling. This design enables our method to preserve semantic skill–expert alignment, mitigate data sparsity, and generate diverse yet contextually coherent teams. Extensive experiments on DBLP and DOTA2 datasets show that our model consistently outperforms state-of-the-art baselines, achieving over 3 × higher recall (16.4% vs. 5.0%) and MAP (9.7% vs. 2.2%) on DBLP, while delivering more than 5 × improvement in MRR (13.3% vs. 2.5%) on DOTA2. Fairness analysis further demonstrates that our method reduces average overlap with the top-100 most popular experts to 2.6, compared to 86.7 for the strongest baseline, and achieves near-optimal diversity with NDKL ≈ 0.1 under high non-popular expert ratios. For reproducibility purposes , we made our code and model publicly available at https://github.com/17shiraz/DiffTF .

The varying performance of information retrieval (IR) methods, including state-of-the-art transformer-based neural retrievers, across diverse queries poses a significant challenge for achieving robust and reliable retrieval effectiveness. Query Performance Prediction (QPP) seeks to estimate the effectiveness of a retrieval method for individual queries, enabling adaptive strategies to improve retrieval outcomes, particularly for challenging queries. However, existing QPP approaches face fundamental challenges: pre-retrieval methods often rely on surface-level query features that fail to capture the nuanced relationship between queries and retrieval effectiveness, while post-retrieval methods depend heavily on the quality of retrieved documents, which can be unreliable for difficult queries. To this end, we propose the Query Space Distance-Based QPP ( QSD-QPP ) framework, which leverages the deterministic and consistent behavior of retrieval methods to estimate query performance by referencing historical queries with known effectiveness. The approach is motivated by the observation that semantically or syntactically similar queries often exhibit consistent retrieval performance, a property that can be exploited to make reliable predictions for unseen queries. QSD-QPP operates in two modes: (1) a lightweight pre-retrieval instantiation that dynamically constructs a query subspace based on embedding distances to interpolate the performance of proximate historical queries, and (2) an enriched post-retrieval instantiation that incorporates contextualized embeddings, document interactions, and historical query associations to enhance prediction accuracy. By utilizing large-scale contextualized embeddings derived from pre-trained language models, QSD-QPP efficiently identifies semantically similar queries and leverages their performance for robust predictions. By addressing the inherent limitations of prior approaches, QSD-QPP achieves a balanced trade-off between computational efficiency, prediction accuracy, and scalability. We evaluate QSD-QPP on four benchmark datasets, including MS MARCO Dev and TREC Deep Learning tracks (2019, 2020, and DL-Hard), demonstrating its superior accuracy and robustness compared to state-of-the-art baselines in both pre-retrieval and post-retrieval QPP tasks. To ensure reproducibility and encourage further research, we publicly release the implementation of our work.

This study examines how music listening is associated with short-term stress expression among individuals with mental health conditions, including depression, anxiety, PTSD, and bipolar disorder, using large-scale social media data. We analyzed over 20 million posts from 10,264 users on Twitter (now X) and identified music-listening sessions through shared links to streaming platforms. Stress-related language was measured in tweets posted within 30–60 min after each listening event. To reduce confounding, we applied Propensity Score Matching (PSM) and modeled associations using Zero-Inflated Generalized Linear Mixed Models (ZIGLMM) across music genres and audio attributes (valence, tempo, instrumentalness). The results reveal clear mental health–specific and genre-dependent effects. For example, users with depression showed 24% higher stress 60 min after listening to pop music, while PTSD users exhibited a 33% increase after 30 min. Low-valence music was associated with delayed stress increases (e.g., 14% in depression and approximately 25% in bipolar), whereas high-valence music showed no significant stress elevation. Building on these findings, we demonstrate a proof-of-concept, stress-aware music recommendation framework that more effectively ranks stress-reducing songs (MRR = 0.35 vs. 0.18 for the base model). These findings highlight the potential of data-driven music interventions for emotional well-being.

Despite recent advances, Large Language Models (LLMs) remain vulnerable to jailbreak attacks that bypass alignment safeguards and elicit harmful outputs. While prior research has proposed various attack strategies differing in human readability and transferability, little attention has been paid to the linguistic and psychological mechanisms that may influence a model's susceptibility to such attacks. In this paper, we examine an interdisciplinary line of research that leverages foundational theories of persuasion from the social sciences to craft adversarial prompts capable of circumventing alignment constraints in LLMs. Drawing on well-established persuasive strategies, we hypothesize that LLMs, having been trained on large-scale human-generated text, may respond more compliantly to prompts with persuasive structures. Furthermore, we investigate whether LLMs themselves exhibit distinct persuasive fingerprints that emerge in their jailbreak responses. Empirical evaluations across multiple aligned LLMs reveal that persuasion-aware prompts significantly bypass safeguards, demonstrating their potential to induce jailbreak behaviors. This work underscores the importance of cross-disciplinary insight in addressing the evolving challenges of LLM safety. The code and data are available. https://github.com/CyberScienceLab/Our-Papers/tree/main/PersuasiveJailbreaking/.

Neural rankers have achieved strong retrieval effectiveness but require large amounts of labeled data, limiting their applicability in few-shot settings. In this paper, we address the sample inefficiency of neural ranking methods by introducing a Reinforcement Learning (RL)-based re-ranking model that achieves high effectiveness with minimal training data. Built on a Deep Q-learning Network (DQN) framework, our approach is designed for few-shot settings, maximizing sample efficiency to ensure robust generalization from limited interactions. Extensive experiments show that our model significantly outperforms data-intensive methods and existing few-shot baselines, demonstrating RL's potential to enhance IR capabilities in few-shot scenarios.

Online hate speech has become a critical issue, particularly during the COVID-19 pandemic, when anti-Asian sentiment surged across social media platforms. However, the causal mechanisms driving emotional and behavioral shifts in users posting hateful content remain understudied. This study investigates the causal relationship between engaging in hateful content and changes in linguistic and emotional expression on social media. Using a dataset of 6,002 Twitter/X users, we employ causal inference techniques, including propensity score matching, and advanced topic modeling to compare users posting hateful content with a matched group of non-hateful users. Our main findings can be summarized as follows: (a) Users who post hateful content show significantly higher levels of anger, anxiety, and negative emotions, along with increased third-person pronoun usage. (b) Moral outrage and profanity levels peak during hateful posts but decline over time, while remaining elevated compared to non-hateful posts. (c) Hateful posts are more interconnected, cover more diverse topics, and are more similar to one another, revealing lower cohesion within individual posts but higher cohesion across posts. These findings contribute to understanding the causal effects of online hate speech on user behavior, offering actionable insights for social media platforms to mitigate the spread of hateful content and its broader societal impact. • Causal inference reveals emotional and linguistic shifts in 6,002 hate speech users. • Hate speech users show heightened anger, anxiety, and fewer positive expressions. • Increased third-person pronouns indicate greater social detachment in hate speech. • Moral outrage and profanity decline over time but stay above control group levels. • Hate speech narratives form cohesive networks with high global cohesion, low specificity.

Factoid entity questions (FEQ), which seek answers in the form of a single entity from knowledge sources, such as DBpedia and Wikidata, constitute a substantial portion of user queries in search engines. This article introduces the knowledge graph embedding model for FEQ (KGE-FEQ) answering. Leveraging a textual knowledge graph derived from extensive text collections, KGE-FEQ encodes textual relationships between entities. The model employs a two-step process: (1) Triple Retrieval, where relevant triples are retrieved from the textual knowledge graph based on semantic similarities to the question, and (2) Answer Selection, where a knowledge graph embedding approach is utilized for answering the question. This involves positioning the embedding for the answer entity close to the embedding of the question entity, incorporating a vector representing the question and textual relations between entities. Extensive experiments evaluate the performance of the proposed approach, comparing KGE-FEQ to state-of-the-art baselines in FEQ answering and the most advanced open-domain question answering techniques applied to FEQs. The results show that KGE-FEQ outperforms existing methods across different datasets. Ablation studies highlights the effectiveness of KGE-FEQ when both the question and textual relations between entities are considered for answering questions.

Predicting users’ interests on social networks is gaining attention due to its potential to cater customized information and services to the end users. Although previous works have extensively explored how users’ interests can be modeled on social networks, there has been limited investigation into the prediction of users’ future interests. The objective of our work in this paper is to empirically study the effectiveness of different sets of features based on users’ past social interactions, historical interests and their temporal dynamics to predict their interests over a collection of future-yet-unobserved topics. More specifically, we introduce and formalize the features for interest prediction in four categories: user-based, topical, explicit user-topic engagement, and friends’ influence. We further explore the influence of temporality by augmenting features with information pertaining to users’ historical interests and social connections. We model the task of future interest prediction as a learning-to-rank problem where different features and their related categories are ranked based on their relevance and performance in interest prediction, and investigate the efficiency of different features individually and comparatively for predicting the future interest of users with different activity levels in social networks over on unobserved topics. After conducting experiments on a real-world dataset sourced from Twitter, we have identified several noteworthy findings: (1) relevance feature in the category of past explicit user-topic engagement is the strongest indicator for predicting user’s future interest across all user groups, with an observed 8.57% decrease in NDCG and an 8.95% decrease in MAP when it is removed in the ablation study. (2) the observation of an 8.06% decrease in NDCG and a 7.3% decrease in MAP, when topical features such as popularity, freshness, and coherence are removed in the ablation study, highlights their significance as among the strongest indicators for users’ future interest, particularly for low-active users. (3) although temporal features show a clear positive impact across user groups with varying levels of activity (resulting in a 4.5% decrease in NDCG and a 7.3% decrease in MAP when removed in the ablation study), the temporal topical features do not demonstrate a significant positive effect, and 4) The removal of user-specific characteristics such as influence and personality traits in the ablation study reveals their significant impact in predicting future interest over cold topics, reflected by a 5.49% decrease in NDCG and a 5.72% decrease in MAP. Our findings make significant contributions to the field of future interest prediction, offering valuable insights and practical implications for various applications in social network analysis.

For the fifth edition of the workshop on Mining Actionable Insights from Social Networks (MAISoN), we organized a special edition with focus on dis/misinformation mining from social media, co-located with CIKM 2020. This topic has attracted a lot of interest from the community since the Coronavirus (COVID-19) epidemic has given rise to an increase of misinformation on social media. The aim of this edition was to bring together researchers from different disciplines interested in mining dis/misinformation on social media. In particular, the distinguishing focus of this special edition was its emphasis on techniques that use social media data for building diagnostic, predictive and prescriptive analysis models related to misinformation. This means that there is rigorous attention for techniques that can be used to understand how and why dis/misinformation is created and spread, to uncover hidden and unexpected aspects of dis/misinformation content, and to recommend insightful countermeasures to restrict the circulation of dis/misinformation and alleviate their negative effects.

In this paper, we implement and publicly share a configurable software workflow and a collection of gold standard datasets for training and evaluating supervised query refinement methods. Existing datasets such as AOL and MS MARCO, which have been extensively used in the literature for this purpose, are based on the weak assumption that users' input queries improve gradually within a search session, i.e., the last query where the user ends her information seeking session is the best reconstructed version of her initial query. In practice, such an assumption is not necessarily accurate for a variety of reasons, e.g., topic drift. The objective of our work is to enable researchers to build gold standard query refinement datasets without having to rely on such weak assumptions. Our software workflow, which generates such gold standard query datasets, takes three inputs: (1) a dataset of queries along with their associated relevance judgements (e.g. TREC topics), (2) an information retrieval method (e.g., BM25), and (3) an evaluation metric (e.g., MAP), and outputs a gold standard dataset. The produced gold standard dataset includes a list of revised queries for each query in the input dataset, each of which effectively improves the performance of the specified retrieval method in terms of the desirable evaluation metric. Since our workflow can be used to generate gold standard datasets for any input query set, in this paper, we have generated and publicly shared gold standard datasets for TREC queries associated with Robust04, Gov2, ClueWeb09, and ClueWeb12. The source code of our software workflow, the generated gold datasets, and benchmark results for three state-of-the-art supervised query refinement methods over these datasets are made publicly available for reproducibility purposes.

Social media users readily share their preferences, life events, sentiment and opinions, and implicitly signal their thoughts, feelings, and psychological behavior. This makes social media a viable source of information to accurately and effectively mine users' interests with the hopes of enabling more effective user engagement, better quality delivery of appropriate services and higher user satisfaction. In this tutorial, we cover five important aspects related to the effective mining of user interests: (1) the foundations of social user interest modeling, such as information sources, various types of representation models and temporal features, (2) techniques that have been adopted or proposed for mining user interests, (3) different evaluation methodologies and benchmark datasets, (4) different applications that have been taking advantage of user interest mining from social media platforms, and (5) existing challenges, open research questions and exciting opportunities for further work.

The focus of point-of-interest recommendation techniques is to suggest a venue to a given user that would match the users' interests and is likely to be adopted by the user. Given the multitude of venues and the sparsity of user check-ins, the problem of recommending venues has shown to be a difficult task. Existing literature has already explored various types of features such as geographical distribution, social structure and temporal behavioral patterns to make a recommendation. In this paper, we propose a new set of features derived based on the neural embeddings of venues and users. We show how the neural embeddings for users and venues can be jointly learnt based on the prior check-in sequence of users and then be used to define three types of features, namely user, venue, and user-venue interaction features. These features are integrated into a feature-based matrix factorization model. Our experiments show that the features defined over the user and venue embeddings are effective for venue recommendation.

Specificity is the level of detail at which a given term is represented. Existing approaches to estimating term specificity are primarily dependent on corpus-level frequency statistics. In this work, we explore how neural embeddings can be used to define corpus-independent specificity metrics. Particularly, we propose to measure term specificity based on the distribution of terms in the neighborhood of the given term in the embedding space. The intuition is that a term that is surrounded by other terms in the embedding space is more likely to be specific while a term surrounded by less closely related terms is more likely to be generic. On this basis, we leverage geometric properties between embedded terms to define three groups of metrics: (1) neighborhood-based, (2) graph-based and (3) cluster-based metrics. Moreover, we employ learning-to-rank techniques to estimate term specificity in a supervised approach by employing the three proposed groups of metrics. We curate and publicly share a test collection of term specificity measurements defined based on Wikipedia's category hierarchy. We report on our experiments through metric performance comparison, ablation study and comparison against the state-of-the-art baselines.

he abundance of user generated content on social networks pro-vides the opportunity to build models that are able to accurately and effectively extract, mine and predict users' interests with the hopes of enabling more effective user engagement, better quality delivery of appropriate services and higher user satisfaction. While traditional methods for building user profiles relied on AI-based preference elicitation techniques that could have been considered to be intrusive and undesirable by the users, more recent advances are focused on a non-intrusive yet accurate way of determining users' interests and preferences. In this tutorial, we cover five important aspects related to the effective mining of user interests: (1) we introduce the information sources that are used for extracting user interests, (2) various types of user interest profiles that have been proposed in the literature, (3) techniques that have been adopted or proposed for mining user interests, (4) the scalability and re-source requirements of the state of the art methods, and finally (5)the evaluation methodologies that are adopted in the literature for validating the appropriateness of the mined user interest profiles.We also introduce existing challenges, open research question and exciting opportunities for further work.

The abundance of user generated content on social networks provides the opportunity to build models that are able to accurately and effectively extract, mine and predict users' interests with the hopes of enabling more effective user engagement, better quality delivery of appropriate services and higher user satisfaction. While traditional methods for building user profiles relied on AI-based preference elicitation techniques that could have been considered to be intrusive and undesirable by the users, more recent advances are focused on a non-intrusive yet accurate way of determining users' interests and preferences. In this tutorial, we cover five important aspects related to the effective mining of user interests: (1) we introduce the information sources that are used for extracting user interests, (2) various types of user interest profiles that have been proposed in the literature, (3) techniques that have been adopted or proposed for mining user interests, (4) the scalability and resource requirements of the state of the art methods, and finally (5) the evaluation methodologies that are adopted in the literature for validating the appropriateness of the mined user interest profiles. We also introduce existing challenges, open research question and exciting opportunities for further work.

The accurate prediction of users' future topics of interests on social networks can facilitate content recommendation and platform engagement. However, researchers have found that future interest prediction, especially on social networks such as Twitter, is quite challenging due to the rapid changes in community topics and evolution of user interactions. In this context, temporal collaborative filtering methods have already been used to perform user interest prediction, which benefit from similar user behavioral patterns over time to predict how a user's interests might evolve in the future. In this paper, we propose that instead of considering the whole user base within a collaborative filtering framework to predict user interests, it is possible to much more accurately predict such interests by only considering the behavioral patterns of the most influential users related to the user of interest. We model influence as a form of causal dependency between users. To this end, we employ the concept of Granger causality to identify causal dependencies. We show through extensive experimentation that the consideration of only one causally dependent user leads to much more accurate prediction of users' future interests in a host of measures including ranking and rating accuracy metrics.

Researchers have shown that it is possible to identify reported instances of personal life events from users' social content, e.g., tweets. This is known as personal life event detection. In this paper, we take a step forward and explore the possibility of predicting users' next personal life event based solely on the their historically reported personal life events, a task which we refer to as personal life event prediction. We present a framework for modeling streaming social content for the purpose of personal life event prediction and describe how various instantiations of the framework can be developed to build a life event prediction model. In our extensive experiments, we find that (i) historical personal life events of a user have strong predictive power for determining the user's future life event; (ii) the consideration of sequence in historically reported personal life events shows inferior performance compared to models that do not consider sequence, and (iii) the number of historical life events and the length of the past time intervals that are taken into account for making life event predictions can impact prediction performance whereby more recent life events show more relevance for the prediction of future life events.