Self-supervised learning (SSL) has become the de facto training paradigm of large models, where pre-training is followed by supervised fine-tuning using domain-specific data and labels. Despite demonstrating comparable performance with supervised methods, comprehensive efforts to assess SSL's impact on machine learning fairness (i.e., performing equally on different demographic breakdowns) are lacking. Hypothesizing that SSL models would learn more generic, hence less biased representations, this study explores the impact of pre-training and fine-tuning strategies on fairness. We introduce a fairness assessment framework for SSL, comprising five stages: defining dataset requirements, pre-training, fine-tuning with gradual unfreezing, assessing representation similarity conditioned on demographics, and establishing domain-specific evaluation processes. We evaluate our method's generalizability on three real-world human-centric datasets (i.e., MIMIC, MESA, and GLOBEM) by systematically comparing hundreds of SSL and fine-tuned models on various dimensions spanning from the intermediate representations to appropriate evaluation metrics. Our findings demonstrate that SSL can significantly improve model fairness, while maintaining performance on par with supervised methods-exhibiting up to a 30% increase in fairness with minimal loss in performance through self-supervision. We posit that such differences can be attributed to representation dissimilarities found between the best- and the worst-performing demographics across models-up to x13 greater for protected attributes with larger performance discrepancies between segments. Code: https://github.com/Nokia-Bell-Labs/SSLfairness

Ubiquitous devices, such as smartphones and wearables, are becoming increasingly popular for monitoring user behavior, health, and well-being. Through omnipresent monitoring, they aim to raise user awareness and encourage positive health behavior change. Yet, ubiquitous technologies suffer from expectation mismatch, lack of user-centric adaptiveness, and, ultimately, high abandonment rates. This work is motivated by the vital need to tailor and personalize ubiquitous technologies while dealing with the challenges arising from the lack of user profiling and the absence of relevant, self-reported user data. To this end, we show that the automatically passively collected sensing data from the wearables can be exploited to improve personalization and infer several user states relevant to demographic, physiological, psychological, and personality aspects, complementing the need for time-consuming self-reports. To accomplish this task, and enable the reproducibility and extensibility of our work; we propose an extensive benchmark suite by exploiting sensing data harvested from ubiquitous devices.Our benchmark covers a wide range of personalization tasks, including modeling gender, age, personality states, and stress, experimenting on the publicly available, newly released LifeSnaps dataset containing over 71 million rows of data capturing the daily lives of 71 participants in their naturalistic environments. The proposed benchmarking continuum showcases the strong potential for the applicability of the presented work in critical applications, such as mental healthcare monitoring, privacy preservation, and responsible artificial intelligence (AI), by fostering fairness assessments when protected attribute knowledge is unavailable.

Online social networks are actively involved in removing malicious social bots due to their role in spreading low-quality information. However, most of the existing bot detectors are supervised classifiers incapable of capturing the evolving behavior of sophisticated bots. Here we propose MulBot, an unsupervised bot detector based on multivariate time series (MTS). For the first time, we exploit multidimensional temporal features extracted from user timelines. We manage the multidimensionality with an LSTM autoencoder, which projects the MTS in a suitable latent space. Then, we perform a clustering step on this encoded representation to identify dense groups of very similar users – a known sign of automation. Finally, we perform a binary classification task achieving f1-score =0.99, outperforming state-of-the-art methods (f1-score ≤ 0.97). Not only does MulBot achieve excellent results in the binary classification task, but we also demonstrate its strengths in a novel and practically-relevant task: detecting and separating different botnets. In this multiclass classification task we achieve f1-score =0.96. We conclude by estimating the importance of the different features used in our model and by evaluating MulBot’s capability to generalize to new unseen bots, thus proposing a solution to the generalization deficiencies of supervised bot detectors.

Modern techniques in intrusion and DoS (Denial of Service) detection tend to be either supervised or semi-supervised, i.e., they require training and labelled data. In this work, we study the problem of correlating security attacks with anomalies reported at runtime by a fully unsupervised outlier detection module, i.e., a component that does not require any training at all. Through a concrete proof-of-concept case study, we demonstrate that unsupervised anomaly detection is both efficient and effective, but still, it needs to be combined with additional mechanisms to yield a complete intrusion detection and prevention solution.

Users in Online Social Networks (OSNs,) leave traces that reflect their personality characteristics. The study of these traces is important for several fields, such as social science, psychology, marketing, and others. Despite a marked increase in research on personality prediction based on online behavior, the focus has been heavily on individual personality traits, and by doing so, largely neglects relational facets of personality. This study aims to address this gap by providing a prediction model for holistic personality profiling in OSNs that includes socio-relational traits (attachment orientations) in combination with standard personality traits. Specifically, we first designed a feature engineering methodology that extracts a wide range of features (accounting for behavior, language, and emotions) from the OSN accounts of users. Subsequently, we designed a machine learning model that predicts trait scores of users based on the extracted features. The proposed model architecture is inspired by characteristics embedded in psychology; i.e, it utilizes interrelations among personality facets and leads to increased accuracy in comparison with other state-of-the-art approaches. To demonstrate the usefulness of this approach, we applied our model on two datasets, namely regular OSN users and opinion leaders on social media, and contrast both samples’ psychological profiles. Our findings demonstrate that the two groups can be clearly separated by focusing on both Big Five personality traits and attachment orientations. The presented research provides a promising avenue for future research on OSN user characterization and classification.

Aggression in online social networks has been studied mostly from the perspective of machine learning, which detects such behavior in a static context. However, the way aggression diffuses in the network has received little attention as it embeds modeling challenges. In fact, modeling how aggression propagates from one user to another is an important research topic, since it can enable effective aggression monitoring, especially in media platforms, which up to now apply simplistic user blocking techniques. In this article, we address aggression propagation modeling and minimization in Twitter, since it is a popular microblogging platform at which aggression had several onsets. We propose various methods building on two well-known diffusion models, Independent Cascade ( IC ) and Linear Threshold ( LT ), to study the aggression evolution in the social network. We experimentally investigate how well each method can model aggression propagation using real Twitter data, while varying parameters, such as seed users selection, graph edge weighting, users’ activation timing, and so on. It is found that the best performing strategies are the ones to select seed users with a degree-based approach, weigh user edges based on their social circles’ overlaps, and activate users according to their aggression levels. We further employ the best performing models to predict which ordinary real users could become aggressive (and vice versa) in the future, and achieve up to AUC = 0.89 in this prediction task. Finally, we investigate aggression minimization by launching competitive cascades to “inform” and “heal” aggressors. We show that IC and LT models can be used in aggression minimization, providing less intrusive alternatives to the blocking techniques currently employed by Twitter.

Abstract Due to their confidence and dominance, narcissistic leaders oftentimes can be perceived favorably by followers, in particular during times of uncertainty. In this study, we propose and examine the relationship between narcissistic leaders and followers who are prone to experience uncertainty intensely and frequently in general, namely highly anxious followers. We do so by applying machine learning algorithms to account for personality traits in a large sample of leaders and followers on Twitter. We find that highly anxious followers are more likely to interact with narcissistic leaders in general, and male narcissistic leaders in particular. Finally, we also examined these interactions in the context of highly popular leaders and found that as leaders become more popular, they begin to attract less anxious followers, regardless of leader gender. We interpret and discuss these findings in relation to previous work and outline limitations and future research recommendations based on our approach.

Cyberbullying and cyberaggression are increasingly worrisome phenomena affecting people across all demographics. More than half of young social media users worldwide have been exposed to such prolonged and/or coordinated digital harassment. Victims can experience a wide range of emotions, with negative consequences such as embarrassment, depression, isolation from other community members, which embed the risk to lead to even more critical consequences, such as suicide attempts. In this work, we take the first concrete steps to understand the characteristics of abusive behavior in Twitter, one of today’s largest social media platforms. We analyze 1.2 million users and 2.1 million tweets, comparing users participating in discussions around seemingly normal topics like the NBA, to those more likely to be hate-related, such as the Gamergate controversy, or the gender pay inequality at the BBC station. We also explore specific manifestations of abusive behavior, i.e., cyberbullying and cyberaggression, in one of the hate-related communities (Gamergate). We present a robust methodology to distinguish bullies and aggressors from normal Twitter users by considering text, user, and network-based attributes. Using various state-of-the-art machine-learning algorithms, we classify these accounts with over 90% accuracy and AUC. Finally, we discuss the current status of Twitter user accounts marked as abusive by our methodology and study the performance of potential mechanisms that can be used by Twitter to suspend users in the future.

Social media dominance largely affects multi-store brands success potential. How can a brand choose the optimal place to locate its stores, given the social media pulse? Which are the suitable metrics to guide such challenging decisions? This work addresses such crucial problems by a novel location casting approach which extracts and integrates knowledge from open data and social media, providing specific indicators and a systematic pipeline for effective locations casting. Emphasis is placed on how the derived knowledge will assess the particular characteristics of accessibility, interest, and centrality to identify fine grained urban indicators. The proposed pipeline predicts the success potential of a chain store's location, under individual or combined such indicators individually. The experimentation under qualitative tests, indicates that the proposed approach provides reliable estimations of brand's locations suitability, and also outperforms existing similar state-of-the-art approaches.

Highly distributed applications dominate today's software industry posing new challenges for novel software architectures capable of supporting real time processing and analytics. The proposed framework, so called REAXICS, is motivated by the fact that the demand for aggregating current and past big data streams requires new software methodologies, platforms and services. The proposed framework is designed to tackle with data intensive problems in real time environments, via services built dynamically under a fully scalable and elastic Lambda based architecture. REAXICS proposes a multi-layer software platform, based on the lambda architecture paradigm, for aggregating and synchronizing real time and batch processing. The proposed software layers and adaptive components support quality of experience, along with community driven software development. Flexibility and elasticity are targeted by hiding the complexity of bootstrapping and maintaining a multi level architecture, upon which the end user can drive queries over input data streams. REAXICS proposes a flexible and extensible software architecture that can capture users preference at the front-end and adapt the appropriate distributed technologies and processes at the back-end. Such a model enables real time analytics in large-scale data driven cloud-based systems.

In this paper, we present a contribution to IR modeling. We propose an approach that computes on the fly, a Personalized Social Document Representation (PSDR) of each document per user based on his social activities. The PSDRs are used to rank documents with respect to a query. This approach has been intensively evaluated on a large public dataset, showing significant benefits for personalized search.

Social networks drive todays opinions and content diffusion. Large scale, distributed and unpredictable social data streams are produced and such evolving data production offers the ground for the data mining and analysis tasks. Such social data streams embed human reactions and inter-relationships and affective and emotional analysis has become rather important in todays applications. This work highlights the major data structures and methodologies used in evolving social data mining and proceeds to the relevant affective analysis techniques. A particular framework is outlined along with indicative applications which employ evolving social data analysis with emphasis on the seminal criteria of topic, location and time. Such mining and analysis overview is beneficial for various scientific and enterpreneural audiences and communities in the social networking area.

We present a technical demonstration of an online city exploration application that helps users identify interesting spots in a city by use of spatio-temporal analysis and clustering of user contributed photos. Our framework analyzes the spatial distribution of large city-centered collections of user contributed photos at different time scales in order to index the most popular spots of a city in a time-aware manner. Subsequently, the photo sets belonging to the same spatiotemporal context are clustered in order to extract representative photos for each spot. The resulting application enables users to obtain flexible summaries of the most important spots in a city given a temporal slice (time of the day, month, season). The demonstration will be based on a photo dataset covering major European cities.

Getting an overview of a large online social net-work and deciding which communities to join is a challenging task for a new user. We propose a method that maps a large network into a smaller graph with two kinds of nodes: a node of the first kind is representative of a community, a node of the second kind is neighbor to a representative and rejects the semantics of that community. Our approach encompasses a learning and ranking algorithm that derives this smaller graph from the original one, and a visualization algorithm that returns a graph layout to the observer. We report on our results on inspecting the network of a folksonomy.

LDAP directories have proliferated as the appropriate storage framework for various and heterogeneous data sources, operating under a wide range of applications and services. Due to the increased amount and heterogeneity of the LDAP data, there is a requirement for appropriate data organization schemes. The LPAIR & LMERGE (LP-LM) algorithm, presented in this article, is a hierarchical agglomerative structure-based clustering algorithm which can be used for the LDAP directory information tree definition. A thorough study of the algorithm’s performance is provided, which designates its efficiency. Moreover, the Relative Link as an alternative merging criterion is proposed, since as indicated by the experimentation, it can result in more balanced clusters. Finally, the LP and LM Query Engine is presented, which considering the clustering-based LDAP data organization, results in the enhancement of the LDAP server’s performance.

Today's web commercial applications demand more powerful recommendation systems due to the rapid increase in the number of both consumers and available products. Searching for the best algorithm with the highest accuracy and realistic complexity is, most of the time, a very costly process in terms of both time and resources. In this paper we suggest an alternative framework called Hydra which enables the virtual fusion of any and as many currently available recommendation algorithms in such a distributed manner that algorithms' complexities are not summarized but parallelized. Therefore, we utilize the available algorithms and technologies aiming to achieve better accuracy in order to surpass even the most state of the art algorithms. In addition, Hydra can be used to find how algorithms interact with each other in order to estimate the resulting accuracy towards inventing a more precise algorithm diminishing the risk of a failed investment. Hydra can be adjusted and integrated in any recommendation application while it is also open to new functionalities which can be embedded easily and in a transparent manner.

Content distribution networks (CDNs) balance costs and quality in services related to content delivery. Devising an efficient content outsourcing policy is crucial since, based on such policies, CDN providers can provide client-tailored content, improve performance, and result in significant economical gains. Earlier content outsourcing approaches may often prove ineffective since they drive prefetching decisions by assuming knowledge of content popularity statistics, which are not always available and are extremely volatile. This work addresses this issue, by proposing a novel self-adaptive technique under a CDN framework on which outsourced content is identified with no a-priori knowledge of (earlier) request statistics. This is employed by using a structure-based approach identifying coherent clusters of "correlated" Web server content objects, the so-called Web page communities. These communities are the core outsourcing unit and in this paper a detailed simulation experimentation has shown that the proposed technique is robust and effective in reducing user-perceived latency as compared with competing approaches, i.e., two communities-based approaches, Web caching, and non-CDN.

Under social tagging systems, a typical Web 2.0 application, users label digital data sources by using freely chosen textual descriptions (tags). Poor retrieval in the aforementioned systems remains a major problem mostly due to questionable tag validity and tag ambiguity. Earlier clustering techniques have shown limited improvements, since they were based mostly on tag co-occurrences. In this paper, a co-clustering approach is employed, that exploits joint groups of related tags and social data sources, in which both social and semantic aspects of tags are considered simultaneously. Experimental results demonstrate the efficiency and the beneficial outcome of the proposed approach in correlating relevant tags and resources.

Web users' clustering is a crucial task for mining information related to users' needs and preferences. Up to now, popular clustering approaches build clusters based on usage patterns derived from users' page preferences. This paper emphasizes the need to discover similarities in users' accessing behavior with respect to the time locality of their navigational acts. In this context, we present two time-aware clustering approaches for tuning and binding the page and time visiting criteria. The two tracks of the proposed algorithms define clusters with users that show similar visiting behavior at the same time period, by varying the priority given to page or time visiting. The proposed algorithms are evaluated using both synthetic and real data sets and the experimentation has shown that the new clustering schemes result in enriched clusters compared to those created by the conventional non-time-aware user clustering approaches. These clusters contain users exhibiting similar access behavior in terms not only of their page preferences but also of their access time.

Web caching and content replication techniques emerged to solve performance problems related to the Web. We propose a generic non-parametric heuristic method that integrates both techniques under a CDN. We provide experimentation showing that our method outperforms the so far separate implementations of Web caching and content replication. Moreover, we show that the performance improvement compared with an existing algorithm is significant. We test all these techniques in a simulation environment under a flash crowd event and a workload of a typical light-weighted CDN operation

In our project, we are adopting a new approach for handling video data. We view the video as a well-defined data type with its own description, parameters and applicable methods. The system is based on PREDATOR, an open-source object-relational DBMS. PREDATOR uses Shore as the underlying storage manager. Supporting video operations (storing, searching-by-content and streaming) and new query types (query-by-example and multi-feature similarity searching) requires major changes in many of the traditional system components. More specifically, the storage and buffer manager has to deal with huge volumes of data with real-time constraints. Query processing has to consider the video methods and operators in generating, optimizing and executing the query plans.