Pub Date : 2025-09-01Epub Date: 2025-07-26DOI: 10.1016/j.osnem.2025.100325
Omran Berjawi , Giuseppe Fenza , Rida Khatoun , Vincenzo Loia
Recommender systems play a crucial role in enhancing user experiences by suggesting content based on users consumption histories. However, a significant challenge they encounter is managing the radicalized contents spreading and preventing users from becoming trapped in radicalized pathways. This paper address the radicalization problem in recommendation systems (RS) by proposing a graph-based approach called Deep Reinforcement Learning Graph Rewiring (DRLGR). First, we measure the radicalization score (Rad(G)) for the recommendation graph by assessing the extent of users’ exposure to radical content. Second, we develop a Reinforcement Learning (RL) method, which learns over time which edges among many possible ones should be rewired to reduce the Rad(G). The experimental results on video and news recommendation datasets show that DRLGR consistently reduces the radicalization score and demonstrates more sustained improvements over time, particularly in more complex graphs compared to baseline methods and heuristic approach such as HEU that may reduce radicalization more rapidly in the early stages with fewer interventions but plateau over time.
{"title":"Mitigating radicalization in recommender systems by rewiring graph with deep reinforcement learning","authors":"Omran Berjawi , Giuseppe Fenza , Rida Khatoun , Vincenzo Loia","doi":"10.1016/j.osnem.2025.100325","DOIUrl":"10.1016/j.osnem.2025.100325","url":null,"abstract":"<div><div>Recommender systems play a crucial role in enhancing user experiences by suggesting content based on users consumption histories. However, a significant challenge they encounter is managing the radicalized contents spreading and preventing users from becoming trapped in radicalized pathways. This paper address the radicalization problem in recommendation systems (RS) by proposing a graph-based approach called Deep Reinforcement Learning Graph Rewiring (DRLGR). First, we measure the radicalization score (Rad(G)) for the recommendation graph by assessing the extent of users’ exposure to radical content. Second, we develop a Reinforcement Learning (RL) method, which learns over time which edges among many possible ones should be rewired to reduce the Rad(G). The experimental results on video and news recommendation datasets show that DRLGR consistently reduces the radicalization score and demonstrates more sustained improvements over time, particularly in more complex graphs compared to baseline methods and heuristic approach such as HEU that may reduce radicalization more rapidly in the early stages with fewer interventions but plateau over time.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"48 ","pages":"Article 100325"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144711820","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-09-01Epub Date: 2025-06-04DOI: 10.1016/j.osnem.2025.100318
Mohammad Majid Akhtar , Navid Shadman Bhuiyan , Rahat Masood , Muhammad Ikram , Salil S. Kanhere
The detection of automated accounts, also known as “social bots”, has been an important concern for online social networks (OSNs). While several methods have been proposed for detecting social bots, significant research gaps remain. First, current models exhibit limitations in detecting sophisticated bots that aim to mimic genuine OSN users. Second, these methods often rely on simplistic profile features, which are susceptible to adversarial manipulation. In addition, these models lack generalizability, resulting in subpar performance when trained on one dataset and tested on another.
To address these challenges, we propose a framework for social Bot detection with Self-Supervised Contrastive Learning (BotSSCL). Our framework leverages contrastive learning to distinguish between social bots and humans in the embedding space to improve linear separability. The high-level representations derived by BotSSCL enhance its resilience to variations in data distribution and ensure generalizability. We evaluate BotSSCL’s robustness against adversarial attempts to manipulate bot accounts to evade detection. Experiments on two datasets featuring sophisticated bots demonstrate that BotSSCL outperforms other supervised, unsupervised, and self-supervised baseline methods. We achieve and higher (F1) performance than SOTA on both datasets. In addition, BotSSCL also achieves 67% F1 when trained on one dataset and tested with another, demonstrating its generalizability under cross-botnet evaluation. Lastly, under adversarial evasion attack, BotSSCL shows increased complexity for the adversary and only allows 4% success to the adversary in evading detection. The code is available at https://github.com/code4thispaper/BotSSCL.
{"title":"BotSSCL: Social Bot Detection with Self-Supervised Contrastive Learning","authors":"Mohammad Majid Akhtar , Navid Shadman Bhuiyan , Rahat Masood , Muhammad Ikram , Salil S. Kanhere","doi":"10.1016/j.osnem.2025.100318","DOIUrl":"10.1016/j.osnem.2025.100318","url":null,"abstract":"<div><div>The detection of automated accounts, also known as “social bots”, has been an important concern for online social networks (OSNs). While several methods have been proposed for detecting social bots, significant research gaps remain. First, current models exhibit limitations in detecting sophisticated bots that aim to mimic genuine OSN users. Second, these methods often rely on simplistic profile features, which are susceptible to adversarial manipulation. In addition, these models lack generalizability, resulting in subpar performance when trained on one dataset and tested on another.</div><div>To address these challenges, we propose a framework for social <strong>Bot</strong> detection with <strong>S</strong>elf-<strong>S</strong>upervised <strong>C</strong>ontrastive <strong>L</strong>earning (BotSSCL). Our framework leverages contrastive learning to distinguish between social bots and humans in the embedding space to improve linear separability. The high-level representations derived by BotSSCL enhance its resilience to variations in data distribution and ensure generalizability. We evaluate BotSSCL’s robustness against adversarial attempts to manipulate bot accounts to evade detection. Experiments on two datasets featuring sophisticated bots demonstrate that BotSSCL outperforms other supervised, unsupervised, and self-supervised baseline methods. We achieve <span><math><mrow><mo>≈</mo><mn>6</mn><mtext>%</mtext></mrow></math></span> and <span><math><mrow><mo>≈</mo><mn>8</mn><mtext>%</mtext></mrow></math></span> higher (F1) performance than SOTA on both datasets. In addition, BotSSCL also achieves 67% F1 when trained on one dataset and tested with another, demonstrating its generalizability under cross-botnet evaluation. Lastly, under adversarial evasion attack, BotSSCL shows increased complexity for the adversary and only allows 4% success to the adversary in evading detection. The code is available at <span><span>https://github.com/code4thispaper/BotSSCL</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"48 ","pages":"Article 100318"},"PeriodicalIF":0.0,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144203492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01Epub Date: 2025-05-10DOI: 10.1016/j.osnem.2025.100313
Imane Khaouja , Daniel Toribio-Flórez , Ricky Green , Cassidy Rowden , Chee Siang Ang , Karen M. Douglas
Social media has become an influential channel for political communication, offering broad reach while enabling the proliferation of misinformation and conspiracy theories. These unchecked conspiracy narratives may allow manipulation by malign actors, posing dangers to democratic processes. Despite their intuitive appeal, little research has examined the strategic usage and timing of conspiracy theories in politicians’ social media communication compared to the spread of misinformation and fake news.
This study provides an empirical analysis of how members of the U.S. Congress spread conspiracy theories on Twitter. Leveraging the Twitter Historical API, we collected a corpus of tweets from members of the US Congress between January 2012 and December 2022. We developed a classifier to identify conspiracy theory content within this political discourse. We also analyzed the linguistic characteristics, topics and distribution of conspiracy tweets. To assess classifier performance, we created ground truth data through human annotation in which experts labeled a sample of 2500 politicians’ tweets.
Our findings shed light on several aspects, including the influence of prevailing political power dynamics on the propagation of conspiracy theories and higher user engagement. Moreover, we identified specific psycho-linguistic attributes within the tweets, characterized by the use of words related to power and causation, and outgroup language. Our results provide valuable insights into the motivations compelling influential figures to engage in the dissemination of conspiracy narratives in political discourse.
{"title":"Political communication and conspiracy theory sharing on twitter","authors":"Imane Khaouja , Daniel Toribio-Flórez , Ricky Green , Cassidy Rowden , Chee Siang Ang , Karen M. Douglas","doi":"10.1016/j.osnem.2025.100313","DOIUrl":"10.1016/j.osnem.2025.100313","url":null,"abstract":"<div><div>Social media has become an influential channel for political communication, offering broad reach while enabling the proliferation of misinformation and conspiracy theories. These unchecked conspiracy narratives may allow manipulation by malign actors, posing dangers to democratic processes. Despite their intuitive appeal, little research has examined the strategic usage and timing of conspiracy theories in politicians’ social media communication compared to the spread of misinformation and fake news.</div><div>This study provides an empirical analysis of how members of the U.S. Congress spread conspiracy theories on Twitter. Leveraging the Twitter Historical API, we collected a corpus of tweets from members of the US Congress between January 2012 and December 2022. We developed a classifier to identify conspiracy theory content within this political discourse. We also analyzed the linguistic characteristics, topics and distribution of conspiracy tweets. To assess classifier performance, we created ground truth data through human annotation in which experts labeled a sample of 2500 politicians’ tweets.</div><div>Our findings shed light on several aspects, including the influence of prevailing political power dynamics on the propagation of conspiracy theories and higher user engagement. Moreover, we identified specific psycho-linguistic attributes within the tweets, characterized by the use of words related to power and causation, and outgroup language. Our results provide valuable insights into the motivations compelling influential figures to engage in the dissemination of conspiracy narratives in political discourse.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"47 ","pages":"Article 100313"},"PeriodicalIF":0.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143928802","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01Epub Date: 2025-05-16DOI: 10.1016/j.osnem.2025.100314
Davide Antonio Mura , Marco Usai , Andrea Loddo, Manuela Sanguinetti, Luca Zedda, Cecilia Di Ruberto, Maurizio Atzori
In recent years, the proliferation of fake news has posed significant challenges to information integrity and public trust, paving the way for the development of artificial intelligence-based models that can analyze information and determine its veracity. This study comprehensively evaluates the Themis architecture in the context of fake news detection on two distinct public datasets: Fakeddit and ReCoVery. To enhance model performance, we systematically investigate various customizations of Themis, including the integration of Low-Rank Adaptation, diverse data augmentation techniques, and multiple configurations, employing the TinyLlama Large Language Model and CLIP ViT image encoders while tuning key parameters to optimize results. Our findings reveal that while the standard Themis model performed adequately, significant improvements were observed by incorporating LoRA and specific data augmentation strategies, particularly in the ReCoVery dataset. Comparisons with existing literature indicate that Themis achieves competitive performance, especially in the ReCoVery dataset, where it outperforms existing solutions.
{"title":"Is it fake or not? A comprehensive approach for multimodal fake news detection","authors":"Davide Antonio Mura , Marco Usai , Andrea Loddo, Manuela Sanguinetti, Luca Zedda, Cecilia Di Ruberto, Maurizio Atzori","doi":"10.1016/j.osnem.2025.100314","DOIUrl":"10.1016/j.osnem.2025.100314","url":null,"abstract":"<div><div>In recent years, the proliferation of fake news has posed significant challenges to information integrity and public trust, paving the way for the development of artificial intelligence-based models that can analyze information and determine its veracity. This study comprehensively evaluates the Themis architecture in the context of fake news detection on two distinct public datasets: Fakeddit and ReCoVery. To enhance model performance, we systematically investigate various customizations of Themis, including the integration of Low-Rank Adaptation, diverse data augmentation techniques, and multiple configurations, employing the TinyLlama Large Language Model and CLIP ViT image encoders while tuning key parameters to optimize results. Our findings reveal that while the standard Themis model performed adequately, significant improvements were observed by incorporating LoRA and specific data augmentation strategies, particularly in the ReCoVery dataset. Comparisons with existing literature indicate that Themis achieves competitive performance, especially in the ReCoVery dataset, where it outperforms existing solutions.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"47 ","pages":"Article 100314"},"PeriodicalIF":0.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144070284","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01Epub Date: 2025-05-17DOI: 10.1016/j.osnem.2025.100315
Paras Stefanopoulos , Sourin Chatterjee , Ahad N. Zehmakan
This paper explores recommender systems in social networks which leverage information such as item rating, intra-item similarities, and trust graph. We demonstrate that item-rating information is more influential than other information types in a collaborative filtering approach. The trust graph-based approaches were found to be more robust to network adversarial attacks due to hard-to-manipulate trust structures. Intra-item information, although sub-optimal in isolation, enhances the consistency of predictions and lower-end performance when fused with other information forms. Additionally, the Weighted Average framework is introduced, enabling the construction of recommendation systems around any user-to-user similarity metric.
{"title":"A first principles approach to trust-based recommendation systems in social networks","authors":"Paras Stefanopoulos , Sourin Chatterjee , Ahad N. Zehmakan","doi":"10.1016/j.osnem.2025.100315","DOIUrl":"10.1016/j.osnem.2025.100315","url":null,"abstract":"<div><div>This paper explores recommender systems in social networks which leverage information such as item rating, intra-item similarities, and trust graph. We demonstrate that item-rating information is more influential than other information types in a collaborative filtering approach. The trust graph-based approaches were found to be more robust to network adversarial attacks due to hard-to-manipulate trust structures. Intra-item information, although sub-optimal in isolation, enhances the consistency of predictions and lower-end performance when fused with other information forms. Additionally, the Weighted Average framework is introduced, enabling the construction of recommendation systems around any user-to-user similarity metric.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"47 ","pages":"Article 100315"},"PeriodicalIF":0.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144070285","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01Epub Date: 2025-05-30DOI: 10.1016/j.osnem.2025.100317
Delfina S. Martinez Pandiani , Erik Tjong Kim Sang , Davide Ceolin
Internet memes are multimodal, highly shareable cultural units that condense complex messages into compact forms of communication, making them a powerful vehicle for information spread. Increasingly, they are used to propagate hateful, extremist, or otherwise ‘toxic’ narratives, symbols, and messages. Research on computational methods for meme toxicity analysis has expanded significantly over the past five years. However, existing surveys cover only studies published until 2022, resulting in inconsistent terminology and overlooked trends. This survey bridges that gap by systematically reviewing content-based computational approaches to toxic meme analysis, incorporating key developments up to early 2024. Using the PRISMA methodology, we extend the scope of prior analyses, resulting in a threefold increase in the number of reviewed works. This study makes four key contributions. First, we expand the coverage of computational research on toxic memes, reviewing 158 content-based studies, including 119 newly analyzed papers, and identifying over 30 datasets while examining their labeling methodologies. Second, we address the lack of clear definitions of meme toxicity in computational research by introducing a new taxonomy that categorizes different toxicity types, providing a more structured foundation for future studies. Third, we observe that existing content-based studies implicitly focus on three key dimensions of meme toxicity—target, intent, and conveyance tactics. We formalize this perspective by introducing a structured framework that models how these dimensions are computationally analyzed across studies. Finally, we examine emerging trends and challenges, including advancements in cross-modal reasoning, the integration of expert and cultural knowledge, the increasing demand for automatic toxicity explanations, the challenges of handling meme toxicity in low-resource languages, and the rising role of generative AI in both analyzing and generating ‘toxic’ memes.
{"title":"‘Toxic’ memes: A survey of computational perspectives on the detection and explanation of meme toxicities","authors":"Delfina S. Martinez Pandiani , Erik Tjong Kim Sang , Davide Ceolin","doi":"10.1016/j.osnem.2025.100317","DOIUrl":"10.1016/j.osnem.2025.100317","url":null,"abstract":"<div><div>Internet memes are multimodal, highly shareable cultural units that condense complex messages into compact forms of communication, making them a powerful vehicle for information spread. Increasingly, they are used to propagate hateful, extremist, or otherwise ‘toxic’ narratives, symbols, and messages. Research on computational methods for meme toxicity analysis has expanded significantly over the past five years. However, existing surveys cover only studies published until 2022, resulting in inconsistent terminology and overlooked trends. This survey bridges that gap by systematically reviewing content-based computational approaches to toxic meme analysis, incorporating key developments up to early 2024. Using the PRISMA methodology, we extend the scope of prior analyses, resulting in a threefold increase in the number of reviewed works. This study makes four key contributions. First, we expand the coverage of computational research on toxic memes, reviewing 158 content-based studies, including 119 newly analyzed papers, and identifying over 30 datasets while examining their labeling methodologies. Second, we address the lack of clear definitions of meme toxicity in computational research by introducing a new taxonomy that categorizes different toxicity types, providing a more structured foundation for future studies. Third, we observe that existing content-based studies implicitly focus on three key dimensions of meme toxicity—target, intent, and conveyance tactics. We formalize this perspective by introducing a structured framework that models how these dimensions are computationally analyzed across studies. Finally, we examine emerging trends and challenges, including advancements in cross-modal reasoning, the integration of expert and cultural knowledge, the increasing demand for automatic toxicity explanations, the challenges of handling meme toxicity in low-resource languages, and the rising role of generative AI in both analyzing and generating ‘toxic’ memes.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"47 ","pages":"Article 100317"},"PeriodicalIF":0.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144167494","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01Epub Date: 2025-05-22DOI: 10.1016/j.osnem.2025.100316
Mehmet Deniz Türkmen , Mucahid Kutlu
As the order of sentences can impact the meaning of texts, transformer models and recurrent neural networks (RNN) also consider the order of the tokens. However, this feature can negatively affect the classification of social media accounts, as users might share messages on entirely different topics in consecutive order. In this study, we explore how to enhance the performance of models that take into account word order for various author profiling tasks on social media. We first draw attention to the transformer models’ input limit and propose a message selection method that also reduces noise caused by irrelevant messages. In addition, we show that arbitrarily concatenating messages can be problematic. Therefore, we propose creating multiple variants of data by shuffling messages, classifying each variant separately, and then aggregating the predictions. In our comprehensive experiments, we focus on age, gender, occupation, and bot detection tasks. We show that the proposed content selection and shuffling-based methods lead to slight improvements in the transformer model’s performance for age and gender detection tasks. However, our approach yields noticeable performance increases for BiLSTM model. Additionally, we observe that the shuffling method serves as an effective means to augment training data, further enhancing models’ performance. Moreover, our shuffling-based approach enhances the models’ resistance to adversarial attacks in gender and occupation detection tasks without compromising their performance in age detection.
{"title":"Message order matters: A robust author profiling approach for social media platforms","authors":"Mehmet Deniz Türkmen , Mucahid Kutlu","doi":"10.1016/j.osnem.2025.100316","DOIUrl":"10.1016/j.osnem.2025.100316","url":null,"abstract":"<div><div>As the order of sentences can impact the meaning of texts, transformer models and recurrent neural networks (RNN) also consider the order of the tokens. However, this feature can negatively affect the classification of social media accounts, as users might share messages on entirely different topics in consecutive order. In this study, we explore how to enhance the performance of models that take into account word order for various author profiling tasks on social media. We first draw attention to the transformer models’ input limit and propose a message selection method that also reduces noise caused by irrelevant messages. In addition, we show that arbitrarily concatenating messages can be problematic. Therefore, we propose creating multiple variants of data by shuffling messages, classifying each variant separately, and then aggregating the predictions. In our comprehensive experiments, we focus on age, gender, occupation, and bot detection tasks. We show that the proposed content selection and shuffling-based methods lead to slight improvements in the transformer model’s performance for age and gender detection tasks. However, our approach yields noticeable performance increases for BiLSTM model. Additionally, we observe that the shuffling method serves as an effective means to augment training data, further enhancing models’ performance. Moreover, our shuffling-based approach enhances the models’ resistance to adversarial attacks in gender and occupation detection tasks without compromising their performance in age detection.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"47 ","pages":"Article 100316"},"PeriodicalIF":0.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144114716","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Early detection of depression in mental health is crucial for better intervention. Social media has been extensively used to examine users’ behavior, motivating researchers to develop an automatic depression detection model. However, the accuracy and clarity of the reasons behind the detection results still need to be improved. Current research focuses primarily on syntactic and semantic information in user-posted texts, while other aspects of users’ psychological characteristics are often overlooked. Therefore, this study addresses the gap by proposing a novel model integrating personality traits and sentiment polarity patterns into an explainable ensemble model. Specifically, we developed two base learners for the averaged and meta-ensemble learning strategy. The first learner employed the Robustly Optimized BERT Pre-training Approach (RoBERTa). For the second learner, we combined the Random Forest and Bidirectional Long Short-Term Memory (RF-BiLSTM) methods to effectively handle the combination of personality traits and sequential information in sentiment polarity patterns. These additional features are obtained by performing domain adaptation for personality prediction and sentiment analysis using a lexicon-based model. Based on the experimental results, our ensemble model improved depression detection results by leveraging the strengths of each base learner. Our model advanced the state-of-the-art, outperforming existing models with an increase in accuracy and F1-score of 4.14% and 2.99%, respectively. The model successfully enhanced the interpretability of detection results, providing a more comprehensive understanding of the factors underlying depressive symptoms. This research highlights the potential of considering alternative additional features as a promising avenue for enhancing depression detection in social media.
{"title":"An explainable ensemble model for revealing the level of depression in social media by considering personality traits and sentiment polarity pattern","authors":"Gede Aditra Pradnyana , Wiwik Anggraeni , Eko Mulyanto Yuniarno , Mauridhi Hery Purnomo","doi":"10.1016/j.osnem.2025.100307","DOIUrl":"10.1016/j.osnem.2025.100307","url":null,"abstract":"<div><div>Early detection of depression in mental health is crucial for better intervention. Social media has been extensively used to examine users’ behavior, motivating researchers to develop an automatic depression detection model. However, the accuracy and clarity of the reasons behind the detection results still need to be improved. Current research focuses primarily on syntactic and semantic information in user-posted texts, while other aspects of users’ psychological characteristics are often overlooked. Therefore, this study addresses the gap by proposing a novel model integrating personality traits and sentiment polarity patterns into an explainable ensemble model. Specifically, we developed two base learners for the averaged and meta-ensemble learning strategy. The first learner employed the Robustly Optimized BERT Pre-training Approach (RoBERTa). For the second learner, we combined the Random Forest and Bidirectional Long Short-Term Memory (RF-BiLSTM) methods to effectively handle the combination of personality traits and sequential information in sentiment polarity patterns. These additional features are obtained by performing domain adaptation for personality prediction and sentiment analysis using a lexicon-based model. Based on the experimental results, our ensemble model improved depression detection results by leveraging the strengths of each base learner. Our model advanced the state-of-the-art, outperforming existing models with an increase in accuracy and F1-score of 4.14% and 2.99%, respectively. The model successfully enhanced the interpretability of detection results, providing a more comprehensive understanding of the factors underlying depressive symptoms. This research highlights the potential of considering alternative additional features as a promising avenue for enhancing depression detection in social media.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"46 ","pages":"Article 100307"},"PeriodicalIF":0.0,"publicationDate":"2025-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143579435","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-05-01Epub Date: 2025-03-05DOI: 10.1016/j.osnem.2025.100308
Mourad Ellouze , Sonda Rekik , Lamia Hadrich Belguith
The effects of psychological crises are evolving at an astounding rate nowadays, presenting a significant challenge for everyone involved in tracking these disorders. Therefore, we propose in this paper a hybrid approach based on linguistic processing and numerical techniques allowing to: (i) identify the presence of psychological emergencies among social network users by analyzing their textual production, (ii) determine the specific type of emergency case, (iii) elaborate a graph for each type of emergency, reflecting the different dimensions linked to the psychological emergency, allowing for a better diagnosis of the situation and providing an overall view of the crisis type, (iv) combine the separate graphs for each emergency to address the various semantic aspects. The work was accomplished using advanced language model techniques, knowledge graphs and neural network graphs. The combination of these techniques ensures that their advantages are leveraged while overcoming their limitations in terms of result generalization. The evaluation of different parts related to detecting the presence of psychological problems, predicting specific type of emergency cases, and detecting links between knowledge graphs was measured using the F-measure metric. The values derived from this measure, corresponding to the evaluation of these three tasks, are, respectively, 83%, 87% and 80%. For the evaluation of the elaboration of each graph related to specific type of emergency cases, this was accomplished using qualitative metric standards. The results obtained can be considered encouraging given the significant scale of our approach.
{"title":"Management of psychological emergency cases on social media: A hybrid approach combining knowledge graphs and graph neural networks","authors":"Mourad Ellouze , Sonda Rekik , Lamia Hadrich Belguith","doi":"10.1016/j.osnem.2025.100308","DOIUrl":"10.1016/j.osnem.2025.100308","url":null,"abstract":"<div><div>The effects of psychological crises are evolving at an astounding rate nowadays, presenting a significant challenge for everyone involved in tracking these disorders. Therefore, we propose in this paper a hybrid approach based on linguistic processing and numerical techniques allowing to: (i) identify the presence of psychological emergencies among social network users by analyzing their textual production, (ii) determine the specific type of emergency case, (iii) elaborate a graph for each type of emergency, reflecting the different dimensions linked to the psychological emergency, allowing for a better diagnosis of the situation and providing an overall view of the crisis type, (iv) combine the separate graphs for each emergency to address the various semantic aspects. The work was accomplished using advanced language model techniques, knowledge graphs and neural network graphs. The combination of these techniques ensures that their advantages are leveraged while overcoming their limitations in terms of result generalization. The evaluation of different parts related to detecting the presence of psychological problems, predicting specific type of emergency cases, and detecting links between knowledge graphs was measured using the F-measure metric. The values derived from this measure, corresponding to the evaluation of these three tasks, are, respectively, 83%, 87% and 80%. For the evaluation of the elaboration of each graph related to specific type of emergency cases, this was accomplished using qualitative metric standards. The results obtained can be considered encouraging given the significant scale of our approach.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"46 ","pages":"Article 100308"},"PeriodicalIF":0.0,"publicationDate":"2025-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143551494","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-05-01Epub Date: 2025-03-20DOI: 10.1016/j.osnem.2025.100312
Giancarlo Sperlì
In this paper, we design a novel data model for a Multimedia Social Network, that has been modeled as an attribute graph for integrating semantic analysis of multimedia content published by users. It combines features inferred from object detection, image classification, and hashtag analysis in a unified model to characterize a user from different points of view. On top of this model, community detection algorithms have been applied to unveil users’ communities. Hence, we design a framework integrating multimedia features with different community detection approaches (topological, deep learning, representation learning, and game theory-based) to improve detection effectiveness. The proposed framework has been evaluated on a real-world dataset, composed of 4.5 million profiles publishing more than 42 million posts and 1.2 million images, to investigate the impact of different features on both graph-building and community detection tasks. The main findings of the proposed analysis show how combining different sets of features inferred from multimedia content allows to achieve the highest modularity score w.r.t. other configurations although it requires a higher running time for building the underlined network. Specifically, representation and game theory-based algorithms achieve the highest results in terms of Modularity measure by exploiting the semantic and contextual information integrated into the proposed model.
{"title":"Community detection in Multimedia Social Networks using an attributed graph model","authors":"Giancarlo Sperlì","doi":"10.1016/j.osnem.2025.100312","DOIUrl":"10.1016/j.osnem.2025.100312","url":null,"abstract":"<div><div>In this paper, we design a novel data model for a Multimedia Social Network, that has been modeled as an attribute graph for integrating semantic analysis of multimedia content published by users. It combines features inferred from object detection, image classification, and hashtag analysis in a unified model to characterize a user from different points of view. On top of this model, community detection algorithms have been applied to unveil users’ communities. Hence, we design a framework integrating multimedia features with different community detection approaches (topological, deep learning, representation learning, and game theory-based) to improve detection effectiveness. The proposed framework has been evaluated on a real-world dataset, composed of 4.5 million profiles publishing more than 42 million posts and 1.2 million images, to investigate the impact of different features on both graph-building and community detection tasks. The main findings of the proposed analysis show how combining different sets of features inferred from multimedia content allows to achieve the highest modularity score w.r.t. other configurations although it requires a higher running time for building the underlined network. Specifically, representation and game theory-based algorithms achieve the highest results in terms of Modularity measure by exploiting the semantic and contextual information integrated into the proposed model.</div></div>","PeriodicalId":52228,"journal":{"name":"Online Social Networks and Media","volume":"46 ","pages":"Article 100312"},"PeriodicalIF":0.0,"publicationDate":"2025-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143683310","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号