SNS search system to display posts on an interactive map from various perspectives

Presented at HCI International 2025 held on 2025/7/7)

Transcript

1. SNS Search System that Uses a Large Language Model to Display Posts on an Interactive Map from Various Perspectives Tomonari Kamba, Ph. D. Information Networking for Innovation and Design (INIAD), Toyo University HCI International 2025
2. Background ⚫Filter bubbles and echo chambers in SNS ⚫Regulation vs. openness ⚫Deleting fake news vs. defining "fake" → Users need diverse perspectives — but it’s difficult Copyright © 2025 by INIAD, Toyo University 2
3. Related Work (1)：Filter Bubbles and Echo Chambers ⚫ Conover et al. (Indiana Univ. 2011, AAAI Conf.): Political Polarization on Twitter ⚫ 250K tweets from 6 weeks before 2010 US midterms ⚫ Retweets → Clear separation by party ⚫ Mentions → One large mixed cluster Copyright © 2025 by INIAD, Toyo University 3 Retweets Mentions
4. Related Work (2)：Encouraging Awareness ⚫ Opinion Space (UC Berkeley, CHI 2010) ⚫ Users express views on topics using sliders → PCA (Principal Component Analysis) used to create a visual opinion map → Encourages comments and reactions to diverse views ⚫ Used by U.S. State Dept. to gather public opinions (2010) Copyright © 2025 by INIAD, Toyo University 4
5. Related Work (3)：Encouraging Balanced Exposure ⚫Jasim et al. (UMass Amherst, CHI 2022): Supporting Serendipitous Discovery … ⚫Visualized how well users covered both positive and negative reviews ⚫Balanced reading → greater confidence in product choices Copyright © 2025 by INIAD, Toyo University 5
6. Related Work (4): Balancing Personalization and General Interest ⚫ An Interactive Personalized Newspaper on the Web (WWW Conf. 1995) ⚫ Automatically personalized news based on reading behavior ⚫ Slider to adjust balance between personal interest and community interest Copyright © 2025 by INIAD, Toyo University 6
7. Research Challenge ⚫How can we present diverse perspectives on a single topic in social media? ⚫On SNS, it’s easy to like, share, or take a stance. ▪But it’s hard to notice or access viewpoints that differ from your own. → What kind of user interface can help reveal these different perspectives? Copyright © 2025 by INIAD, Toyo University 7
8. Proposed System ⚫SNS search system ⚫When the user enters a query, an interactive map of the results is shown ⚫The system identifies multiple perspectives and places the results accordingly ⚫Users can discover different viewpoints and reflect on each post based on its position Copyright © 2025 by INIAD, Toyo University 8
9. Movie Copyright © 2025 by INIAD, Toyo University 9
10. Copyright © 2025 by INIAD, Toyo University 10
11. SNS search module User：SNS search X (twitter) API SNS search results Classification axes inquiry module GPT-4o (OpenAI API) - SNS search results - Axes inquiry prompt Classification axes Post score inquiry module - Posts and axes - Scores inquiry prompt Scores Map layout Search keyword System Behavior
12. User Interface Copyright © 2025 by INIAD, Toyo University 12 Axes found by the LLM Selection on the map syncs with the post on the list
13. Prompts to the LLM to Identify Axes (partly simplified) ⚫ Analyze the following tweets and propose a set of discussion axes. ⚫ Generate {axis_count} axes. For each axis, provide: ▪ A concise axis name ▪ A name for the negative pole ({pole1}) ▪ A name for the positive pole ({pole2}) ⚫ Choose axes that best reflect the content of the tweets and promote meaningful discussion. ⚫ Use clear and easy-to-understand language. ⚫ Make sure each axis represents a distinct perspective, avoiding redundancy. Copyright © 2025 by INIAD, Toyo University 13
14. Examples of the Identified Axes（The experiment on Aug. 4, 2024) Query Axes Pole 1 (negative) Pole 2 (positive) Global warming Climate Change Perception Denial Acknowledgement Emotional Response Inaction Concern Action on Climate Change Inaction Proactive Measures GPU Sentiment Towards GPU Technology Technical Challenges/Complexities Excitement/Satisfaction Technical Knowledge and Usability Technical Challenges/Complexities Ease of Use/Solutions Impact of GPU on Gaming Experience Negative Impact Positive Impact Generative AI Ethical Implications of AI AI Unethical Concerns AI Ethical Practices AI Impact on Industries Minimal Industry Impact Transformative Industry Impact Community and AI Development Community Ignored Community engaged Copyright © 2025 by INIAD, Toyo University 14
15. Examples of the Scores Query Post text Axis and score Global warming I remember the summers of my childhood, playing outside without care. Now, I worry about the future of our planet. global warming before it's too late. Perception: +1.0 Emotional Response: +0.9 "Global warming" and "climate change" are shams. Perception: -1.0 Emotional Response: -0.8 GPU Get at least a 4070 GPU and an i7 14700K and 64GB RAM. Rest of the components won't really effect performance. Technological Development 0.8 User Experience 0.6 Can't stream a game because my CPU and GPU max out while playing ... this is HELL. ・Technological Development: - 0.7 User Experience Copyright © 2025 by INIAD, Toyo University 15
16. Experiment: Impact on Browsing Behavior (10 university students) Map Condition Search words selected Unique words Posts viewed View per word list map No map 8.1 7.1 65.8 9.3 With map 8.3 5.3 11.1 62.3 13.8 Copyright © 2025 by INIAD, Toyo University 16 ⚫ Each participant used the interface with and without the map (10 minutes each; order randomized) ⚫ For each session, participants were shown 17 queries × 100 posts; search results were displayed instantly after query selection ⚫ With the map, the number of unique search terms decreased, but views per word increased by 50% ⚫ When the map was present, 85% of viewed posts were accessed via the map
17. Map Interaction Patterns and User Impressions Match Copyright © 2025 by INIAD, Toyo University 17 I can view a wide variety of posts according to my interests. I didn't want to see negative ones, so I selectively avoided them. If there's a map, I only browse the ones I'm inter- ested in. Without it, I'd go from the top. I can grasp the contents quickly. I don't have to look at the negative ones． I can pick and choose what I want to see. I can read each post more carefully. I never got tired of looking at the same keywords. I could see both positive and the critical views．
18. Expert Interviews: Open-Ended Use and Impressions ⚫ IT Expert: University professor specializing in CG and VR/AR ▪ Searched topics like GPU and security incidents. ▪ "It’s hard to come up with axes myself." ▪ Requested features to categorize results by company or by type of problem ⚫ Journalist: Experienced reporter in political and economic affairs ▪ Searched topics like the U.S. presidential election and Japan–U.S. relations ▪ Political criticism can come from very different standpoints, and a single axis often doesn’t capture the full picture Copyright © 2025 by INIAD, Toyo University 18
19. Summary ⚫LLMs are effective for extracting viewpoints, not just classifying by sentiment ⚫The interactive map significantly changes user behavior ⚫User browsing tendencies can be inferred from map interaction history Copyright © 2025 by INIAD, Toyo University 19
20. Future Challenges ⚫ Can LLM-based advice lead to long-term behavioral change? ⚫ Should we encourage users to view negative content, even if they prefer to avoid it? ⚫ How can we present non-binary perspectives more effectively? Copyright © 2025 by INIAD, Toyo University 20
21. ⚫ Users’ interest in diverse content can be inferred from their search behavior. ⚫ Frequent clicks on negative X- axis tweets suggest a response to negative tone, but positive Y-axis tweets also attract clicks—indicating content matters too. Copyright © 2025 by INIAD, Toyo University 21 Example of operation history analysis by LLM (partially simplified)