Algorithmic recommendation systems are increasingly shaping children's digital consumption, but there is limited understanding of how audio-visual features impact the visibility and popularity of videos aimed at young audiences. Regulatory frameworks, such as the EU Digital Services Act, demand greater transparency and accountability, particularly regarding content targeted at minors. However, current systems often overlook the influence of content design on engagement. In this position paper, we propose a research agenda to systematically analyze interpretable visual and audio features, such as color vividness, motion intensity, vocal dynamics, and musicality. By linking these elements to engagement outcomes, we aim to discover consistent patterns that can inform the design of child-sensitive recommendation systems, algorithmic audits, and compliance with policies, as well as establish a foundation for more accountable algorithmic media environments for children.
Recommendation systems play a crucial role in shaping the digital media consumption of millions of
users, particularly children, who are increasingly exposed to algorithmically curated content on platforms
such as YouTube [
To address this gap, we propose a new research perspective that systematically analyzes the role of intrinsic audio-visual features in shaping the popularity and virality of children’s videos. We argue that understanding how low-level multimedia signals correlate with engagement is significant to ensure transparent, fair, and child-appropriate recommendation systems. Our position is that such content features, despite their potential psychological and perceptual impact, are critically understudied in both academic research and regulatory scrutiny. Our perspective requires the development of a multimodal analysis pipeline that can extract interpretable audio-visual descriptors from full-length video content. By investigating how these features vary across high-visibility and low-visibility content aimed at children, future research can uncover methodical engagement patterns and evaluate their alignment with child protection principles and legal obligations under frameworks such as the DSA.
Despite interest in video content analysis, especially for child safety and moderation, gaps remain in
understanding how intrinsic audio-visual features influence the visibility and popularity of children’s
videos on platforms like YouTube. Most existing research focuses on high-level content classification
or metadata-driven moderation. Deep learning (DL) models are commonly used to detect
inappropriate content [
We propose to systematically extract low-level audio-visual features from children’s videos to uncover correlations between content characteristics and their popularity or recommendation exposure. We argue that formalizing the influence of content-level attributes can provide valuable insights for more fair and transparent recommenders. In practice, in the initial iteration, we plan to analyze the top-25 most-viewed YouTube videos for children compared with the overall top-25 most-viewed videos.
On the visual side, we plan to compute global color statistics (saturation, brightness, contrast) to capture the overall vividness and visual salience of each video, factors previously shown to afect children’s attention [26]. Texture descriptors, such as local binary patterns, will be used to distinguish visually detailed scenes [27]. Motion characteristics will be quantified using the dense optical flow metric [28], which captures the proportion of highly dynamic frames, an indicator of editing pace and visual stimulation. We also intend to include features that reflect social and narrative structure, such as face presence and wave-ratio, to approximate the visibility of on-screen presenters or characters [29]. To identify diferences in animation style , we compute depth statistics and assess consistency across frames to infer whether the video adopts a flat 2-D layout or 3-D CGI production [30]. To complement these handcrafted features, we plan to extract vector embeddings from pre-trained deep video models such as VideoMAE [31] and SlowFast [32], which can summarize spatial and temporal patterns to describe the video’s overall visual style and movement patterns.
On the audio side, we plan to extract rhythmic and harmonic descriptors, such as spectral contrast, melody, and tempo, to distinguish between speech, music, and song-based content [33, 34]. Additional features, including short-term energy, zero-crossing rate, spectral centroid and roll-of, harmonic ratio, pitch, and silence ratio, will serve as proxies for vocal intensity and excitement level, which have been linked to children’s emotional arousal [35].
The extracted features will be used in comparative analyses between content explicitly targeted at children and general-audience videos to measure whether children’s videos exhibit unique design signatures. In conclusion, this planned methodology seeks to inform both the auditing of content-driven engagement mechanisms and the development of recommender systems that are more accountable and appropriate for child audiences.
The author have not employed any Generative AI tools. Engagement From Features of Time- and Value-Continuous, Dimensional Emotions, Frontiers in Computer Science 4 (2022). [13] A. C. Munaro, R. H. Barcelos, E. C. F. Mafezzolli, J. P. S. Rodrigues, E. C. Paraiso, To engage or not engage? The features of video content on YouTube afecting digital consumer engagement, Journal of Consumer Behaviour 20 (2021) 1336–1352. doi:10.1002/cb.1939. [14] X. Li, M. Shi, X. S. Wang, Video mining: Measuring visual information using automatic methods, International Journal of Research in Marketing 36 (2019) 216–231. doi:10.1016/j.ijresmar. 2019.02.004. [15] Y. Beyhan, M. S. Pera, Covering Covers: Characterization Of Visual Elements Regarding Sleeves, in: Adjunct Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization, UMAP ’23 Adjunct, Association for Computing Machinery, New York, NY, USA, 2023, pp. 28–33. doi:10.1145/3563359.3597404. [16] A. Milton, L. Batista, G. Allen, S. Gao, Y.-K. D. Ng, M. S. Pera, “Don’t Judge a Book by its Cover”: Exploring Book Traits Children Favor, in: Proceedings of the 14th ACM Conference on Recommender Systems, RecSys ’20, Association for Computing Machinery, New York, NY, USA, 2020, pp. 669–674. doi:10.1145/3383313.3418490. [17] L. Spear, A. Milton, G. Allen, A. Raj, M. Green, M. D. Ekstrand, M. S. Pera, Baby Shark to Barracuda: Analyzing Children’s Music Listening Behavior, in: Proceedings of the 15th ACM Conference on Recommender Systems, RecSys ’21, Association for Computing Machinery, New York, NY, USA, 2021, pp. 639–644. doi:10.1145/3460231.3478856. [18] M. M. Neumann, C. Herodotou, Evaluating YouTube videos for young children, Education and
Information Technologies 25 (2020) 4459–4475. doi:10.1007/s10639-020-10183-7. [19] D. Poveda, M. Matsumoto, E. Sundin, H. Sandberg, C. Aliagas, J. Gillen, Space and practices: Engagement of children under 3 with tablets and televisions in homes in Spain, Sweden and England, Journal of Early Childhood Literacy 20 (2020) 500–523. doi:10.1177/1468798420923715. [20] J. Zhang, Y. Huang, M. Gao, Video Features, Engagement, and Patterns of Collective Attention
Allocation: An Open Flow Network Perspective, Journal of Learning Analytics 9 (2022) 32–52. [21] J. Balanzategui, Examining the "Elsagate" Phenomenon: Disturbing Children’s YouTube Content and New Frontiers in Children’s Culture, AoIR Selected Papers of Internet Research (2019). doi:10.5210/spir.v2019i0.10921. [22] W. Han, M. Ansingkar, Discovery of Elsagate: Detection of Sparse Inappropriate Content from Kids Videos, in: 2020 Zooming Innovation in Consumer Technologies Conference (ZINC), 2020, pp. 46–47. doi:10.1109/ZINC50678.2020.9161808. [23] P. Soustas, M. Edwards, The Elsagate Corpus: Characterising Commentary on Alarming Video Content, in: R. Mitkov, S. Ezzini, T. Ranasinghe, I. Ezeani, N. Khallaf, C. Acarturk, M. Bradbury, M. El-Haj, P. Rayson (Eds.), Proceedings of the First International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security, International Conference on Natural Language Processing and Artificial Intelligence for Cyber Security, Lancaster, UK, 2024, pp. 147– 152. [24] S. Lackmann, P.-M. Léger, P. Charland, C. Aubé, J. Talbot, The Influence of Video Format on Engagement and Performance in Online Learning, Brain Sciences 11 (2021) 128. doi:10.3390/ brainsci11020128. [25] S. K. Mridha, B. Sarkar, S. Chatterjee, M. Bhattacharyya, ViSSa: Recognizing the appropriateness of videos on social media with on-demand crowdsourcing, Information Processing & Management 57 (2020) 102189. doi:10.1016/j.ipm.2019.102189. [26] D. R. Anderson, H. L. Kirkorian, Attention and Television, in: Psychology of Entertainment,
Routledge, 2006. [27] T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (2002) 971–987. doi:10.1109/TPAMI.2002.1017623. [28] G. Farnebäck, Two-Frame Motion Estimation Based on Polynomial Expansion, in: J. Bigun, T. Gustavsson (Eds.), Image Analysis, Springer, Berlin, Heidelberg, 2003, pp. 363–370. doi:10. 1007/3-540-45103-X_50. [29] R. F. Kizilcec, J. N. Bailenson, C. J. Gomez, The instructor’s face in video instruction: Evidence from two large-scale field studies, Journal of Educational Psychology 107 (2015) 724–739. doi: 10. 1037/edu0000013. [30] R. Ranftl, A. Bochkovskiy, V. Koltun, Vision Transformers for Dense Prediction, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 12179–12188. [31] Z. Tong, Y. Song, J. Wang, L. Wang, VideoMAE: Masked Autoencoders are Data-Eficient Learners for Self-Supervised Video Pre-Training, Advances in Neural Information Processing Systems 35 (2022) 10078–10093. [32] C. Feichtenhofer, H. Fan, J. Malik, K. He, SlowFast Networks for Video Recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 6202–6211. [33] G. Tzanetakis, P. Cook, Musical genre classification of audio signals, IEEE Transactions on Speech and Audio Processing 10 (2002) 293–302. doi:10.1109/TSA.2002.800560. [34] J. Salamon, E. Gomez, Melody Extraction From Polyphonic Music Signals Using Pitch Contour Characteristics, IEEE Transactions on Audio, Speech, and Language Processing 20 (2012) 1759–1770. doi:10.1109/TASL.2012.2188515. [35] H. E. Kragness, M. J. Eitel, A. M. Baksh, L. J. Trainor, Evidence for early arousal-based diferentiation of emotions in children’s musical production, Developmental Science 24 (2021) e12982. doi:10. 1111/desc.12982.