TrueFluence – Social Media Promotion Video Credibility Analysis

TrueFluence – Social Media Promotion Video Credibility Analysis

Abhijith J Nair, Anupriya A Pillai, Ashwin M, VR Athira, Shabanamol S, Dr. Sabeena K, Chinchu M Pillai

Department of Computer Science and Engineering, College of Engineering Chengannur, Kerala, India

Abstract – The booming social media and e-commerce have turned the verifying of digital content into a major challenge of modern research. Misleading content, including misinformation, counterfeit reviews, and non- genuine online profiles, endangers not only society but also commerce and public discourse. This paper systematically reviews ten contemporary scholarly articles focusing on the evaluation of credibility and authenticity in different online domains such as influencer marketing, fake news, and review accuracy. We engage in the analysis and comparison of the various methods that these studies describe, categorizing them under two primary concepts: feature engi- neering techniques and the employment of machine learning models. Our review reveals an evident change in the field, where the dependence on single data types has been replaced by more complex methodologies that amalgamate several modalities. In the same vein, we see a shift from conventional machine learning algorithms to sophisticated deep learning frameworks like Transformers and Graph Convolutional Networks. A central revelation from this research is that source and user-related features are gaining importance and are getting as significant as content-based features for strong credibility assessments. We draw a comparative framework from these various approaches, recognizing the contributions, advantages, and disadvantages of each study. Ultimately, this paper underscores the significant challenges that remain, especially the requirement for extensive, standardized datasets with multi-level labels, and it points out the major research areas that will help to develop more effective and widely applicable systems for online content verification.

Index Terms – Fake News Detection, Credibility Assessment, Sentiment Analysis, Deep Learning, Multi- modal Analysis, Social Media Analytics, Review Spam Detection.

1. INTRODUCTION

   The current digital world has greatly changed the lifecycle

   of information, from the very beginning of its creation to the consumption of it. Nevertheless, this digital transformation has inadvertently provided a support system for the propagation of false informationthe misleading reviews and disingenu- ous online behavior [9]alongside the empowered individual content creators. This change in the information dissemination process is extensive, thus, the public being misled through fake news [4], [8], and more specific to the marketplace, the consumer being deceived through false product reviews [7], [10] and the influencers image being supported through in- flated engagement metrics [2], [3]. Consequently, the academic and research sectors are facing an extreme challenge which is critical in all aspects: developing automated means that can proficiently judge the authenticity and reliability of digital content as their capability has to be second to none.

   The very first explorations into this field generally devoted their attention to single traits, such as the communication styles found in a document or basic statistics derived from a users account. Nevertheless, the activities of the bad guys are ever-changing which leads to the necessity of developing more sophisticated and comprehensive detection techniques. Todays studies bring this persistent arms race to light, and they show a clear trend to the shifting towards complicated models that are able to assimilate and utilize a wide variety of data signals. The present paper combines and thoroughly assesses ten cutting-edge investigations in this field. The chosen studies address some of the most important issues:

   • Detecting Fake News and Misinformation: The detec- tion process consists of the assessment of news articles as well as their outreach through social media and short video platforms [4], [8].
   • Verifying Online Review Authenticity: This aspect pertains to the differentiation among consumer reviews on e-commerce and other service platforms that are either real or false [7], [10].
   • Evaluating User and Influencer Credibility: This seg- ment identifies the quality of social media accounts by including the verification of an influencers interaction and his/her audiences loyalty [2], [3], [5],
   [6].

   The paper presents a methodical critique of these studies. The different feature engineering strategies categorized into source-based, content-based, and multi-modal are discussed in Section II. The application of machine learning and deep learning models to these features is analyzed in Section III. A thorough comparison of the ten papers is given in Section IV, where their main aims, methods, and discoveries are summarized in a table. Finally, Section V presents the major trends, points out the existing gaps in research, and suggests the areas that hold promise for future studies.

2. FEATURE ENGINEERING FOR CREDIBILITY ASSESSMENT

   The quality and significance features enumerated from the data are the basis upon which any credibility detection sys- tem rests. The examination of the literature here uncovers a plethora of features, which can be reasonably categorized into three major divisions.

   1. User, Source, and Profile-based Features

      One of the most powerful and increasingly popular methods to judge the contents credibility is to check the source of the content. Content has not become the only concern for many modern systems that now check the creator behind it. According to Sitaula et al. [9], for the case of fake news detection, source credibility is a stronger indicator than the content in the first place. Their study points out the number of authors, their publication records (especially past links to fake news), and their co-authorship networks as features. They draw a conclusion that the spreaders of misinformation rarely share the same pool with the authors of truth [9].

      This user-centric approach is echoed in other domains. For instance, Hu et al. [2] introduced the CISER model to quantify reviewer credibility through three innovative factors: trust, expertise, and network influence. In a similar vein, Qi et al. [8] observed that legitimate news publishers on short video platforms tend to have verified accounts and demonstrate more prolific content creation activities (such as more videos and fans) than their counterparts who publish fake news [8]. Moradi and Chehreghani [5] adopted a comprehensive strat- egy, utilizing a broad set of profile features including the accounts creation date and the number of followers. A sub- sequent SHAP analysis in their study pinpointed the account creation time as the most influential predictive feature in their model [5].

   2. Content-based and Textual Features

      Content, in its essence, still plays a leading role in credibility analysis. Deviations in style and language can often hint at

      falsified material. Abedin et al. citeb7 made it a point in their research to examine only content features while disre- garding any information about the reviewer so as to imitate a consumers view. The metrics they used for their model included review length, subjectivity, readability, and rating extremity [7]. A SHAP analysis pointed subjectivity and length as the most important prdictors [7].

   3. Multi-modal and Social Context Features

      Online content these days is very rarely only text, so multi- modal methodologies have been a part of many recent re- searches. These methods bring together materials from differ- ent data types like text, images, audio, and video. Liu et al. [1] suggested a sentiment analysis model that is a combination of visual features from a 3D CNN, textual features, and audio features through a novel Multiple Kernel Learning (MKL) algorithm, thus providing a more complete picture of the

      sentiment shown in a video [1]. Also, Qi et al. [8] formed SV- FEND, a multi-modal framework that helps to find fake news on short video platforms and also deals with video keyframes, audio, and text transcripts [8].

      Yet, the media alone does not speak for itself, and the social context around it provides important hints for the credibility assessment. Kim et al. [3] looked at the engagement patterns on Instagram to judge audience loyalty and authenticity. They created multi-relational graphs from likes and comments and used contextualized embeddings (via Longformer) from user comments to get a better picture of user behavior [3]. In a big quantitative study, Narassiguin and Sargent [6] observed a positive correlation between engagement rates and factors such as the number of hashtags and the presence of certain objects in pictures. Their investigation revealed that posts that had fewer hashtags and those that had people or pets in them received more engagement [6].

3. METHODOLOGIES AND MODELS

   Having been extracted, the features are then supplied to a variety of classification models. The reviewed literature quite plainly reveals a methodological changeover from the hurly- burly of conventional machine learning to the quietness of the state-of-the-art deep learning.

   1. Traditional Machine Learning Approaches

      A few studies are still making use of the classical machine learning models, either as the main classification method or as a point of reference for comparison. For instance, Moon et al. [4] examined a variety of classifiers, among them Random Forest and SVM, and concluded that Random Forest gave the best performance for predicting the dissemination of fake news, attaining an accuracy of 92.05% [4]. Sitaula et al.

      [9] also tried out different classifiers, with Logistic Regression

      and Linear SVM getting the highest F1-scores on their dataset [9]. Those models still attract users because of their excellent interpretability and low processing requirements, particularly when dealing with structured numerical data.

   2. Deep Learning Architectures

      Deep learning has taken over the field of research and become the primary method because of the potential to decipher complex patterns and representations directly from unprocessed data.

      • Convolutional Neural Networks (CNNs): CNNs are a popular choice when it comes to feature extraction from both images and texts. In the study conducted by Abd-Alhalem et al. [10], a CNN is used to get deep features from the review texts. These are then forwarded to be combined with the aspect-based ones [10]. Liu et al. [1] utilize a 3D CNN [1] to get the spatio-temporal information from video.
      • Recurrent Neural Networks (RNNs) and LSTMs: The literature reviews these papers using RNNs as a non- main model, yet they still point out the strong ability of RNNs and LSTMs to treat sequential data, thus making them apt for linguistic pattern text analysis.
      • Transformers and Language Models: The Transformer framework in particular BERT type models has been adopted as the de facto model for natural language pro- cessing tasks. Moradi and Chehreghani [5] apply BERT to the generation of tweet embeddings while Kim et al. [3] resort to Longformer to accommodate lengthy user comment sequences for more profound user behavior analysis [3], [5]. Qi et al. [8] even go a step further and use BERT for just text analysis and at the same time employ cross-modal transformers for information merging across text, audio, and vision streams [8].
        • Graph Convolutional Networks (GCNs): For networks with an inherent structure, such as social interactions, GCNs are very powerful. Kim et al. [3] presented ALAIM, a multi- relational GCN framework that was created to simulate the interactions between influencers and their followers for predicting both retention.
   3. Fusion and Multi-task Frameworks

   A single feature set or model very often is not adequate for a complicated task. As a result, many sophisticated systems use the concepts of information fusion or multi-task learning. Liu and coworkers [1] presented a Multiple Kernel Learning (MKL) algorithm, which integrates features from various modalities and dynamically assigns them weights according to their predictive power [1]. Abd-Alhalem et al. [10] proposed a fusion network, which combines the deep features obtained

   by a CNN with the engineered aspect-based features [10].

   In another method, Kim et al. [3] applied a multi-task learning framework to make concurrent predictions for two related metrics: audience loyalty and authenticity [3]. This approach allows the model to produce user representations that are more robust and generalizable. The MultiCred model of Moradi and Chehreghani [5] is another example of a fusion strategy, as it combines embeddings from profile features, tweets, and comments before the final classification step [5].

4. Comparative Analysis

   In order to present a clear overview of the studies conducted so far, the table I gathers and shows all ten articles by some important dimensions: the main focus of each paper, the methodologies and features used in the work, the datasets employed, and the main contributions. This comparison not only points out the diversity of methods in the area but also reveals some common trends overall.

   The table quite clearly depicts a path leading to advanced complexity and integration. Earlier studies, for example, com- monly made use of features that were hand-crafted and conventional machine learning while the newer ones almost exclusively depended on deep learning to automate representa- tion learning from data that was unstructured and multimodal. There is, moreover, a remarkable change towards the inclusion of social context and network analysis that allows one to see beyond the contents isolated viewpoint.

   The analysis presents a major challenge that can be at- tributed to the very popular practice of using custom, domain- specific datasets such as FakeSV [8] or even data gathered implicitly for a particular study [5]. Consequently, it becomes difficult direct, apples-to-apples to compare model perfor- mance between different studies and this again points at the pressing need for standard public benchmark datasets. On the other side, in the nature of the classification task, we see a significant difference. It is the case that many studies treat credibility as a binary choice (e.g., real vs. fake), whereas the study by Moradi and Chehreghani [5] proposes a more intricate, multi-level assessment of credibility, which is more reflective of its subtle nature [5].

   TABLE I – Comparative Analysis of Recent Studies in Credibility Assessment

   Paper Focus & Data Technical Approach (Features & Methodology)

   Key Contribution

   Liu et al. [1] Focus: Multi-modal Sentiment Features: Visual, textual, and audio. Novel MKL algorithm dynamically Analysis. Data: MOUD, IEMOCAP Method: 3D CNN and Multiple Kernel weights multi-modal features for video datasets. Learning (MKL) for fusion. improved video sentiment analysis.

   Hu et al. [2] Focus: Recommender Systems; Features: Reviewer credibility (trust, CISER model integrates eviewer Reviewer Credibility. Data: Amazon expertise, network); content aesthetics. credibility and user interest for robust Camera Reviews. Method: Heuristic profiling and recommendations.

   fastText sentiment analysis.

   Kim et al. [3] Focus: Influencer Marketing; Audience Features: Social engagement (likes, ALAIM framework uses GCNs to Authenticity. Data: Large-scale comments); comment embeddings. jointly model audience loyalty and Instagram data. Method: Multi-task, Multi-relational authenticity from social interactions.

   GCN; Longformer.

   Moon et al. [4] Focus: Fake News Diffusion. Data: Features: User influence; Content Models fake news diffusion by FibVID (COVID-19 News). complexity; topical interest. Method: combining user influence with their

   T-LDA for topic modeling; Random topical interest alignment. Forest classifier.

   Moradi & Focus: Multi-level User Credibility. Features: Rich profile, tweet, and MultiCred model and new dataset for Chehreghani [5] Data: Custom-collected Twitter data. comment features. Method: DNN with multi-level (non-binary) user credibility

   BERT/DistilBERT embeddings. assessment.

   Narassiguin & Sar- gent [6]

   Focus: Influencer Marketing Analytics. Data: 713k+ influencers across 5 platforms.

   Features: Engagement rate, demographics, hashtags, image objects. Method: Statistical analysis; Image object detection.

   Large-scale empirical study correlating engagement rate with diverse influencer features.

   Abedin et al. [7] Focus: Online Review Credibility.

   Data: Yelp reviews.

   Features: Content-only: Length, subjectivity, readability, consistency. Method: Deep Learning; SHAP for explainability.

   Predicts review credibility from a consumer perspective (no reviewer info) using explainable AI.

   Qi et al. [8] Focus: Fake News on Short Video Platforms. Data: FakeSV (custom Chinese dataset).

   Features: Multi-modal (video, audio, text); Social context. Method:

   Cross-modal Transformers for fusion.

   Contributes FakeSV, the largest short video fake news dataset, and a strong multi-modal baseline.

   Sitaula et al. [9] Focus: Credibility-based Fake News

   Detection. Data: PolitiFact, Buzzfeed.

   Features: Source-based (author info); Content-based (sentiment, typos).

   Method: Logistic Regression, SVM.

   Shows that source credibility features are stronger predictors of fake news than content features.

   Abd-Alhalem et al. [10]

   Focus: Fake Review Detection in

   E-commerce. Data: Amazon Reviews.

   Features: Hybrid of CNN deep features and aspect-based features (PoS, GloVe). Method: CNN and Aspect fusion network.

   Novel fusion of deep and aspect-based features significantly improves fake review detection.

5. Conclusion and Future Directions

In sum, this review has synthesized and contrasted ten latest publications that investigated online credibility evaluation au- tomation. The field has rapidly and distinctly evolved accord- ing to our interpretation. Methodologically, the isolation of feature engineering and the adoption of holistic, multi-modal, and context-aware frameworks mark the trend significantly. Deep learning architectures have risen as the go-to methods, especially Transformers for text processing and GCNs for analyzing network data, because of their ability to extract powerful representations from intricate datasets.

Throughout the feature engineering process, the main re- search interest has migrated from merely looking at what is being said (the content) and who is saying it (the source) to considering how the audience is responding (the social context) as well. The work of Sitaula et al. [9] and Hu et al. [2] provides convincing evidence that the credibility of the source and the reviewer can sometimes be even more powerful predictors than the content itself. This finding implies very strongly that the future systems will need to find ways to integrate these user-centric signals effectively.

Even though there have been considerable improvements, there are still major obstacles in the way. The lack of uniform annotations for large-scale benchmark datasets that are pub- licly available hinders the evaluation of various models side by side and eventually causes the field to move slower. The use of customized datasets, though advantageous for certain cases, is a major contributor to the disjointedness of the research community. Additionally, the majority of studies still regard reliability as a binary classification issue. The multi- level evaluation research by Moradi and Chehreghani [5] is an essential step towards the gradual acceptance of the notion that reliability varies and is not just a flip of the coin.

Research efforts in the future should focus on three signif- icant areas:

1. Creating Benchmark Datasets: The research commu- nity needs to make it a priority to develop massive datasets that are multi-modal and multi-level labeled. These materials are the backbone of strong model development and the guarantee of fair and consistent evaluations across studies.
2. Improving Model Explainability: The more sophisti- cated the models are, particularly those based on deep learning, the more likely their black box characteristic to be a major barrier to acceptance and trust. The need for integrating explainability techniques is very urgent, as practiced by Abedin et al. [7] with their application of SHAP, to decipher model decisions and create trust in the corresponding outputs.
3. Enhancing Generalizability and Robustness: The fu- ture models should be inherently robust against any form of adversarial manipulation and at the same time be able to generalize across different platforms, languages, and

domains. It wont be easy but pioneering studies in that direction are going to rely on such techniques as transfer learning and domain adaptation to make it possible to move beyond platform dependency.

In short, the battle against online fraud and misinformation is a continuous and challenging problem that needs constant innovation. The studies reviewed are revealing a vibrant and flexible sector that is progressively tapping multi-modal data and advanced deep learning methods to develop more efficient and subtle systems for deciding authenticity and trustworthi- ness.

REFERENCES

1. J. Liu, Z. Wang, G. Wan, and J. Liu, A Novel Multi-modal Sentiment Analysis Based on Multiple Kernel Learning with Margin-Dimension Constraint, International Journal of Computational Intelligence Sys- tems, vol. 17, p. 207, 2024.
2. S. Hu, A. Kumar, F. Al-Turjman, S. Gupta, S. Seth, and Shubham, Reviewer Credibility and Sentiment Analysis Based User Profile Mod- elling for Online Product Recommendation, IEEE Access, vol. 8, pp. 26172-26189, 2020.
3. S. Kim, X. Chen, J.-Y. Jiang, J. Han, and W. Wang, Evaluating Au- dience Loyalty and Authenticity in Influencer Marketing via Multi-task Multi-relational Learning, in Proceedings of the Fifteenth International AAAI Conference on Web and Social Media (ICWSM 2021), 2021, pp. 278-289.
4. U. F. Moon, M. A. H. Rasel, and M. M. Anwar, Modeling the Sharing and Diffusion of Fake News in Social Media, in 2024 IEEE Interna- tional Conference on Signal Processing, Information, Communication and Systems (SPICSCON), 2024.
5. M. Moradi and M. H. Chehreghani, Multilevel User Credibility As- sessment in Social Networks, arXiv preprint arXiv:2309.13305, 2025.
6. A. Narassiguin and S. Sargent, Data Science for Influencer Mar- keting: feature processing and quantitative analysis, arXiv preprint arXiv:1906.05911, 2019.
7. E. Abedin, A. Mendoza, P. Akbarighatar, and S. Karunasekera, Pre- dicting Credibility f Online Reviews: An Integrated Approach, IEEE Access, vol. 12, pp. 49050-49061, 2024.
8. P. Qi, Y. Bu, J. Cao, W. Ji, R. Shui, J. Xiao, D. Wang, and T.-

   S. Chua, FakeSV: A Multimodal Benchmark with Rich Social Context for Fake News Detection on Short Video Platforms, arXiv preprint arXiv:2211.10973, 2022.

9. N. Sitaula, C. K. Mohan, J. Grygiel, X. Zhou, and R. Za- farani, Credibility-based Fake News Detection, arXiv preprint arXiv:1911.00643, 2019.
10. S. M. Abd-Alhalem, H. A. Ali, N. F. Soliman, A. D. Algarni, and

H. S. Marie, Advancing E-Commerce Authenticity: A Novel Fusion Approach Based on Deep Learning and Aspect Features for Detecting False Reviews, IEEE Access, vol. 12, pp. 116055-116070, 2024.