204 - Using machine learning to explore adolescent perspectives on consuming social media content

Publication Number: 4201.204

Kayla Kern, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States; Christopher N. Cascio, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States; Megan A. Moreno, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States; Ellen Selkie, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States

Background: Adolescent social media research requires investigation of experiences beyond quantification of screen time. Little is known about the processes adolescents use to consume social media content that meets their individual needs, despite this content’s potential to impact their health and wellbeing.

Objective: This study explores how adolescents consume and perceive social media content using a novel machine learning approach to analyze large amounts of qualitative data.

Design/Methods: Adolescents aged 13 - 15 years from Wisconsin completed 1-hour semi-structured interviews between October 2023 - October 2025. Participants discussed their approach to finding social media content and how content made them feel. Researchers categorized segments from the Linguistic Inquiry and Word Count (LIWC)-22 Contextualizer for thematic analysis of content-seeking approaches. LIWC was also used to assess semantic components (including positive and negative valence) of responses.

Results: Of 289 participants (demographics in Table 1), 255 shared how they access social media content. In thematic analysis of contextualized segments, scrolling was the most popular method of finding content, followed by searching. Participants described scrolling for content they enjoyed, then using other features such as TikTok’s For You Page; Instagram’s Reels, Stories, and Explore and Home Pages; and personal messages, notifications, and saved content to curate feeds. Teens also reported searching creator pages, hashtags, and audio for curation. While some participants described “hop[ing to tell] the algorithm to keep giving [them similar content],” others referenced the process as “building up” their feeds. In terms of sentiment detected by machine learning (Table 2), teens used language that indicated Need, Want, and Acquire (i.e., both active and passive) states related to social media feeds. Motives for content acquisition were described using Curiosity and Allure terms. Of the 87 participants that reflected on emotional responses (Table 3), the majority described content with positive tone and emotion compared to negative tone and emotion.

Conclusion(s): Using a unique method for large-scale qualitative data analysis, we found that adolescents generally consume content on social media through passive exposure (i.e., scrolling) as opposed to active seeking. Adolescents generally spoke positively about social media content they consumed. Future research should assess differences in types of content and emotional response between exposure mechanisms and design implications for adolescent wellbeing.

Table 1. Summary of Participant Demographics
Table 1 (Collapsed).pdf*Participants were able to select more than one option - % is higher than 100.

Table 2. Prevalence of LIWC Dimensions when Describing Approaches to Finding Content
Table 2.pdf*Descriptions are outlined in The Development and Psychometric Properties of LIWC-22

Table 3. Prevalence of Positive and Negative Valence when Describing Accounts
Table 3.pdf*16 participants did not indicate a positive or negative valence