The widespread use of Multimedia Messaging Service (MMS) has led to a significant increase in the circulation of malicious visual content, presenting new challenges for scalable content moderation systems. In this work, we address the problem of visual spam detection in MMS by introducing a domain-specific taxonomy of inappropriate image categories. Based on this taxonomy, we construct a balanced training dataset from publicly available image collections, and two additional evaluation benchmarks derived from real-world MMS messages, which to the best of our knowledge are not covered by existing public datasets. All datasets were verified and manually labeled in order to ensure high annotation quality in line with our taxonomy. Furthermore, we show how to efficiently classify across eight categories related to MMS spam using an adapted CLIP-based architecture. Our empirical evaluation demonstrates that a fine-tuned CLIP model achieves strong accuracy that closely matches the performance of GPT-4o, but at a significantly lower cost which is crucial when performing at scale.

Wikipedia, the Web's largest encyclopedia, frequently faces content disputes or malicious users seeking to subvert its integrity. Administrators can mitigate such disruptions by enforcing "page protection" that selectively limits contributions to specific articles to help prevent the degradation of content. However, this practice contradicts one of Wikipedia's fundamental principles—that it is open to all contributors—and may hinder further improvement of the encyclopedia. In this paper, we examine the effect of page protection on article quality to better understand whether and when page protections are warranted. Using decade-long data on page protections from the English Wikipedia, we conduct a quasi-experimental study analyzing pages that received "requests for page protection"—written appeals submitted by Wikipedia editors to administrators to impose page protections. We match pages that indeed received page protection with similar pages that did not and quantify the causal effect of the interventions on a well-established measure of article quality. Our findings indicate that the effect of page protection on article quality depends on the characteristics of the page prior to the intervention: high-quality articles are affected positively, as opposed to low-quality articles that are impacted negatively. Subsequent analysis suggests that high-quality articles degrade when left unprotected, whereas low-quality articles improve. Overall, with our study, we outline page protections on Wikipedia and inform best practices on whether and when to protect an article.

Point-of-interest (POI) recommender systems help users discover relevant locations, but their effectiveness is often compromised by popularity bias, which disadvantages less popular, yet potentially meaningful places. This paper addresses this challenge by evaluating the effectiveness of context-aware models and calibrated popularity techniques as strategies for mitigating popularity bias. Using four real-world POI datasets (Brightkite, Foursquare, Gowalla, and Yelp), we analyze the individual and combined effects of these approaches on recommendation accuracy and popularity bias. Our results reveal that context-aware models cannot be considered a uniform solution, as the models studied exhibit divergent impacts on accuracy and bias. In contrast, calibration techniques can effectively align recommendation popularity with user preferences, provided there is a careful balance between accuracy and bias mitigation. Notably, the combination of calibration and context-awareness yields recommendations that balance accuracy and close alignment with the users' popularity profiles, i.e., popularity calibration.

Narrative-driven recommenders aim to provide personalized suggestions for user requests expressed in free-form text such as ''I want to watch a thriller with a mind-bending story, like Shutter Island.'' Although large language models (LLMs) have been shown to excel in processing general natural language queries, their effectiveness for handling such recommendation requests remains relatively unexplored. To close this gap, we compare the performance of 38 open- and closed-source LLMs of various sizes, such as LLama 3.2 and GPT-4o, in a movie recommendation setting. For this, we utilize a gold-standard, crowdworker-annotated dataset of posts from reddit's movie suggestion community and employ various prompting strategies, including zero-shot, identity, and few-shot prompting. Our findings demonstrate the ability of LLMs to generate contextually relevant movie recommendations, significantly outperforming other state-of-the-art approaches, such as doc2vec. While we find that closed-source and large-parameterized models generally perform best, medium-sized open-source models remain competitive, being only slightly outperformed by their more computationally expensive counterparts. Furthermore, we observe no significant differences across prompting strategies for most models, underscoring the effectiveness of simple approaches such as zero-shot prompting for narrative-driven recommendations. Overall, this work offers valuable insights for recommender system researchers as well as practitioners aiming to integrate LLMs into real-world recommendation tools.

Abstract The ability to efficiently resolve conflicts in interactive constraint-based applications is critical for user experience and system reliability. Conflict resolution can be regarded as a specific type of explanation, often denoted as diagnosis. Existing work on integrating machine learning with diagnostic reasoning emphasizes on the combination of hitting set approaches with probabilistic reasoning and memory-based machine learning. An alternative to such two-phase diagnosis approaches is direct diagnosis, which focuses on determining diagnoses without predetermining conflicts. In this article, we utilize diagnosis knowledge from the past to improve diagnosis efficiency while also maintaining user-defined preference criteria. Our approach integrates model-based collaborative filtering (feed-forward neural networks) and other machine learning approaches (e.g., logistic regression and random forest) with direct model-based diagnosis ( FastDiag ). The re-ordering of constraints as input to the diagnosis algorithm increases the efficiency of diagnostic reasoning for determining preference-preserving diagnoses. Through experiments on real-world configuration knowledge bases ( B2C , BusyBox , EA and Linux kernel ), we demonstrate significant runtime improvements and high accuracy in diagnosis prediction. With this, we also contribute to the growing body of literature on combining machine learning and constraint-based reasoning.

We present evidence for the existence and effectiveness of adversarial attacks on graph neural networks (GNNs) that aim to degrade fairness. These attacks can disadvantage a particular subgroup of nodes in GNN-based node classification, where nodes of the underlying network have sensitive attributes, such as race or gender. We conduct qualitative and experimental analyses explaining how adversarial link injection impairs the fairness of GNN predictions. For example, an attacker can compromise the fairness of GNN-based node classification by injecting adversarial links between nodes belonging to opposite subgroups and opposite class labels. Our experiments on empirical datasets demonstrate that adversarial fairness attacks can significantly degrade the fairness of GNN predictions (attacks are effective) with a low perturbation rate (attacks are efficient) and without a significant drop in accuracy (attacks are deceptive). This work demonstrates the vulnerability of GNN models to adversarial fairness attacks. We hope our findings raise awareness about this issue in our community and lay a foundation for the future development of GNN models that are more robust to such attacks.

Abstract Complex systems, abstractly represented as networks, are ubiquitous in everyday life. Analyzing and understanding these systems requires, among others, tools for community detection. As no single best community detection algorithm can exist, robustness across a wide variety of problem settings is desirable. In this work, we present Synwalk, a random walk-based community detection method. Synwalk builds upon a solid theoretical basis and detects communities by synthesizing the random walk induced by the given network from a class of candidate random walks. We thoroughly validate the effectiveness of our approach on synthetic and empirical networks, respectively, and compare Synwalk’s performance with the performance of Infomap and Walktrap (also random walk-based), Louvain (based on modularity maximization) and stochastic block model inference. Our results indicate that Synwalk performs robustly on networks with varying mixing parameters and degree distributions. We outperform Infomap on networks with high mixing parameter, and Infomap and Walktrap on networks with many small communities and low average degree. Our work has a potential to inspire further development of community detection via synthesis of random walks and we provide concrete ideas for future research.

In this paper, we quantify the impact of self- and cross-excitation on the temporal development of user activity in Stack Exchange Question & Answer (Q&A) communities. We study differences in user excitation between growing and declining Stack Exchange communities, and between those dedicated to STEM and humanities topics by leveraging Hawkes processes. We find that growing communities exhibit early stage, high cross-excitation by a small core of power users reacting to the community as a whole, and strong long-term self-excitation in general and cross-excitation by casual users in particular, suggesting community openness towards less active users. Further, we observe that communities in the humanities exhibit long-term power user cross-excitation, whereas in STEM communities activity is more evenly distributed towards casual user self-excitation. We validate our findings via permutation tests and quantify the impact of these excitation effects with a range of prediction experiments. Our work enables researchers to quantitatively assess the evolution and activity potential of Q&A communities.

Ontologies in the biomedical domain are numerous, highly specialized and very expensive to develop. Thus, a crucial prerequisite for ontology adoption and reuse is effective support for exploring and finding existing ontologies. Towards that goal, the National Center for Biomedical Ontology (NCBO) has developed BioPortal---an online repository containing more than 500 biomedical ontologies. In 2016, BioPortal represents one of the largest portals for exploration of semantic biomedical vocabularies and terminologies, which is used by many researchers and practitioners. While usage of this portal is high, we know very little about how exactly users search and explore ontologies and what kind of usage patterns or user groups exist in the first place. Deeper insights into user behavior on such portals can provide valuable information to devise strategies for a better support of users in exploring and finding existing ontologies, and thereby enable better ontology reuse. To that end, we study and group users according to their browsing behavior on BioPortal and use data mining techniques to characterize and compare exploration strategies across ontologies. In particular, we were able to identify seven distinct browsing types, all relying on different functionality provided by BioPortal. For example, Search Explorers extensively use the search functionality while Ontology Tree Explorers mainly rely on the class hierarchy for exploring ontologies. Further, we show that specific characteristics of ontologies influence the way users explore and interact with the website. Our results may guide the development of more user-oriented systems for ontology exploration on the Web.

When users interact with the Web today, they leave sequential digital trails on a massive scale. Examples of such human trails include Web navigation, sequences of online restaurant reviews, or online music play lists. Understanding the factors that drive the production of these trails can be useful, for example, for improving underlying network structures, predicting user clicks, or enhancing recommendations. In this work, we present a method called HypTrails for comparing a set of hypotheses about human trails on the Web, where hypotheses represent beliefs about transitions between states. Our method utilizes Markov chain models with Bayesian inference. The main idea is to incorporate hypotheses as informative Dirichlet priors and to calculate the evidence of the data under them. For eliciting Dirichlet priors from hypotheses, we present an adaption of the so-called (trial) roulette method, and to compare the relative plausibility of hypotheses, we employ Bayes factors. We demonstrate the general mechanics and applicability of HypTrails by performing experiments with (i) synthetic trails for which we control the mechanisms that have produced them and (ii) empirical trails stemming from different domains including Web site navigation, business reviews, and online music played. Our work expands the repertoire of methods available for studying human trails.

We present a new method for detecting interpretable subgroups with exceptional transition behavior in sequential data. Identifying such patterns has many potential applications, e.g., for studying human mobility or analyzing the behavior of internet users. To tackle this task, we employ exceptional model mining, which is a general approach for identifying interpretable data subsets that exhibit unusual interactions between a set of target attributes with respect to a certain model class. Although exceptional model mining provides a well-suited framework for our problem, previously investigated model classes cannot capture transition behavior. To that end, we introduce first-order Markov chains as a novel model class for exceptional model mining and present a new interestingness measure that quantifies the exceptionality of transition subgroups. The measure compares the distance between the Markov transition matrix of a subgroup and the respective matrix of the entire data with the distance of random dataset samples. In addition, our method can be adapted to find subgroups that match or contradict given transition hypotheses. We demonstrate that our method is consistently able to recover subgroups with exceptional transition models from synthetic data and illustrate its potential in two application examples. Our work is relevant for researchers and practitioners interested in detecting exceptional transition behavior in sequential data.

Ever since the inception of the Web website administrators have tried to steer user browsing behavior for a variety of reasons. For example, to be able to provide the most relevant information, for offering specific products, or to increase revenue from advertisements. One common approach to steer or bias the browsing behavior of users is to influence the link selection process by, for example, highlighting or repositioning links on a website. In this paper, we present a methodology for (i) expressing such navigational biases based on the random surfer model, and for (ii) measuring the consequences of the implemented biases. By adopting a model-based approach we are able to perform a wide range of experiments on seven empirical datasets. Our analyses allows us to gain novel insights into the consequences of navigational biases. Further, we unveil that navigational biases may have significant effects on the browsing processes of users and their typical whereabouts on a website. The first contribution of our work is the formalization of an approach to analyze consequences of navigational biases on the browsing dynamics and visit probabilities of specific pages of a website. Second, we apply this approach to analyze several empirical datasets and improve our understanding of the effects of different biases on real-world websites. In particular, we find that on webgraphs - contrary to undirected networks - typical biases always increase the certainty of the random surfer when selecting a link. Further, we observe significant side effects of biases, which indicate that for practical settings website administrators might need to carefully balance the desired outcomes against undesirable side effects.

Many online collaboration networks struggle to gain user activity and become self-sustaining due to the ramp-up problem or dwindling activity within the system. Prominent examples include online encyclopedias such as (Semantic) MediaWikis, Question and Answering portals such as StackOverflow, and many others. Only a small fraction of these systems manage to reach self-sustaining activity, a level of activity that prevents the system from reverting to a nonactive state. In this article, we model and analyze activity dynamics in synthetic and empirical collaboration networks. Our approach is based on two opposing and well-studied principles: (i) without incentives, users tend to lose interest to contribute and thus, systems become inactive, and (ii) people are susceptible to actions taken by their peers (social or peer influence). With the activity dynamics model that we introduce in this article we can represent typical situations of such collaboration networks. For example, activity in a collaborative network, without external impulses or investments, will vanish over time, eventually rendering the system inactive. However, by appropriately manipulating the activity dynamics and/or the underlying collaboration networks, we can jump-start a previously inactive system and advance it toward an active state. To be able to do so, we first describe our model and its underlying mechanisms. We then provide illustrative examples of empirical datasets and characterize the barrier that has to be breached by a system before it can become self-sustaining in terms of critical mass and activity dynamics. Additionally, we expand on this empirical illustration and introduce a new metricp—theActivity Momentum—to assess the activity robustness of collaboration networks.

In this paper, we analyze the influence of social status on opinion dynamics and consensus building in collaboration networks. To that end, we simulate the diffusion of opinions in empirical collaboration networks by taking into account both the network structure and the individual differences of people reflected through their social status. For our simulations, we adapt a well-known Naming Game model and extend it with the Probabilistic Meeting Rule to account for the social status of individuals participating in a meeting. This mechanism is sufficiently flexible and allows us to model various situations in collaboration networks, such as the emergence or disappearance of social classes. In this work, we concentrate on studying three well-known forms of class society: egalitarian, ranked and stratified. In particular, we are interested in the way these society forms facilitate opinion diffusion. Our experimental findings reveal that (i) opinion dynamics in collaboration networks is indeed affected by the individuals' social status and (ii) this effect is intricate and non-obvious. In particular, although the social status favors consensus building, relying on it too strongly can slow down the opinion diffusion, indicating that there is a specific setting for each collaboration network in which social status optimally benefits the consensus building process.

When users interact with the Web today, they leave sequential digital trails on a massive scale. Examples of such human trails include Web navigation, sequences of online restaurant reviews, or online music play lists. Understanding the factors that drive the production of these trails can be useful for e.g., improving underlying network structures, predicting user clicks or enhancing recommendations. In this work, we present a general approach called HypTrails for comparing a set of hypotheses about human trails on the Web, where hypotheses represent beliefs about transitions between states. Our approach utilizes Markov chain models with Bayesian inference. The main idea is to incorporate hypotheses as informative Dirichlet priors and to leverage the sensitivity of Bayes factors on the prior for comparing hypotheses with each other. For eliciting Dirichlet priors from hypotheses, we present an adaption of the so-called (trial) roulette method. We demonstrate the general mechanics and applicability of HypTrails by performing experiments with (i) synthetic trails for which we control the mechanisms that have produced them and (ii) empirical trails stemming from different domains including website navigation, business reviews and online music played. Our work expands the repertoire of methods available for studying human trails on the Web.

Algorithms for constructing hierarchical structures from user-generated metadata have caught the interest of the academic community in recent years. In social tagging systems, the output of these algorithms is usually referred to as folksonomies (from folk-generated taxonomies). Evaluation of folksonomies and folksonomy induction algorithms is a challenging issue complicated by the lack of golden standards, lack of comprehensive methods and tools as well as a lack of research and empirical/simulation studies applying these methods. In this article, we report results from a broad comparative study of state-of-the-art folksonomy induction algorithms that we have applied and evaluated in the context of five social tagging systems. In addition to adopting semantic evaluation techniques, we present and adopt a new technique that can be used to evaluate the usefulness of folksonomies for navigation . Our work sheds new light on the properties and characteristics of state-of-the-art folksonomy induction algorithms and introduces a new pragmatic approach to folksonomy evaluation, while at the same time identifying some important limitations and challenges of folksonomy evaluation. Our results show that folksonomy induction algorithms specifically developed to capture intuitions of social tagging systems outperform traditional hierarchical clustering techniques. To the best of our knowledge, this work represents the largest and most comprehensive evaluation study of state-of-the-art folksonomy induction algorithms to date.

Today, a number of algorithms exist for constructing tag hierarchies from social tagging data. While these algorithms were designed with ontological goals in mind, we know very little about their properties from an information retrieval perspective, such as whether these tag hierarchies support efficient navigation in social tagging systems. The aim of this paper is to investigate the usefulness of such tag hierarchies (sometimes also called folksonomies - from folk-generated taxonomy) as directories that aid navigation in social tagging systems. To this end, we simulate navigation of directories as decentralized search on a network of tags using Kleinberg's model. In this model, a tag hierarchy can be applied as background knowledge for decentralized search. By constraining the visibility of nodes in the directories we aim to mimic typical constraints imposed by a practical user interface (UI), such as limiting the number of displayed subcategories or related categories. Our experiments on five different social tagging datasets show that existing tag hierarchy algorithms can support navigation in theory, but our results also demonstrate that they face tremendous challenges when user interface (UI) restrictions are taken into account. Based on this observation, we introduce a new algorithm that constructs efficiently navigable directories on our datasets. The results are relevant for engineers and scientists aiming to improve navigability of social tagging systems.

Recently, a number of algorithms have been proposed to obtain hierarchical structures - so-called folksonomies - from social tagging data. Work on these algorithms is in part driven by a belief that folksonomies are useful for tasks such as: (a) Navigating social tagging systems and (b) Acquiring semantic relationships between tags. While the promises and pitfalls of the latter have been studied to some extent, we know very little about the extent to which folksonomies are pragmatically useful for navigating social tagging systems. This paper sets out to address this gap by presenting and applying a pragmatic framework for evaluating folksonomies. We model exploratory navigation of a tagging system as decentralized search on a network of tags. Evaluation is based on the fact that the performance of a decentralized search algorithm depends on the quality of the background knowledge used. The key idea of our approach is to use hierarchical structures learned by folksonomy algorithm as background knowledge for decentralized search. Utilizing decentralized search on tag networks in combination with different folksonomies as hierarchical background knowledge allows us to evaluate navigational tasks in social tagging systems. Our experiments with four state-of-the-art folksonomy algorithms on five different social tagging datasets reveal that existing folksonomy algorithms exhibit significant, previously undiscovered, differences with regard to their utility for navigation. Our results are relevant for engineers aiming to improve navigability of social tagging systems and for scientists aiming to evaluate different folksonomy algorithms from a pragmatic perspective.