Source-linked AI summary
A Pedagogically Demonstrative Model Visualizing the Pathway from Online Interactions to Personalized Recommendation
Sushmita Khan, Connor Pennington, Bart P Knijnenburg
TL;DR
Many users have not been taught how digital activity shapes personalized recommendations, a process that can be difficult to understand. This paper presents an educational artifact that simulates the process with an LLM and Sankey visualization. In a pilot study, participants reported better conceptual understanding and privacy awareness, but anticipated little behavior change because platform participation felt unavoidable.
Problem
Many users have not been taught how their digital activities shape personalized recommendations, despite the growing role of personalization in online experiences.
Source: §1 · p. 1Method
Results
The pilot evaluation found that the illustrative model helped young adults understand how online activities shape their experiences and prompted privacy awareness, but participants did not expect to change behavior.
Source: §1 · pp. 1–2, §5 · p. 6, §6 · p. 7Takeaways & Limitations
The artifact can support comprehension and privacy reflection among young adults, while participants’ perceived dependence on digital platforms constrained anticipated behavior change.
Source: §5 · p. 6, §6 · p. 7The six-person pilot used another person’s data, and the simplified artifact prioritizes saliency over accuracy, so some representations could mislead.
Source: §5 · pp. 6–7
Abstract
from arXiv · showhide
Personal digital activity increasingly shapes online experiences, yet few users have been educated regarding the processes transforming raw interactions into personalized suggestions. We developed an education artifact that illustratively simulates how AI leverages users' digital activities to shape online recommendations (e.g., ads). Our artifact processes users' digital activity using a locally-hosted LLM to generate user profiles of their inferred interests and personalized recommendations. A three-layered Sankey diagram maps data sources through inferred interests to personalized recommendations. Interactive filters enable users to explore how different combinations of data sources influence personalized outcomes. This paper describes the artifact and its educational value, and reports findings of a pilot think-aloud study with six young adults. We find that the artifact effectively taught participants the conceptual relationship between digital activities and personalized recommendations. While this did lead participants to develop privacy awareness, they anticipated minimal behavior change due to the perceived unavoidability of platform participation.
1. Introduction and Related Work
Many users have not learned how their digital activity shapes personalized experiences, despite the potential risks for young adults. The paper introduces a visual educational artifact and reports that it helps participants understand personalization, while they expect little behavioral change.
1. Introduction and Related Work
- Limited understanding of how digital activity shapes personalization is especially concerning for young adults facing vulnerability to manipulation, discrimination, and reduced autonomy.Prior work cited in the passage reports that tangible presentations, such as visual representations, help younger adults understand algorithmic decision-making.Source: §1 · p. 1
- The artifact processes users’ digital activities with a locally hosted LLM and visualizes data sources, inferred interests, and recommendations in a three-layered Sankey diagram.The study asks whether this diagram helps users conceptualize how digital activities shape personalized recommendations, such as online ads.Source: §1 · p. 1
- Participants developed privacy awareness but did not anticipate changing their behavior because they viewed platform participation as difficult to avoid.The passage also notes that most users were unwilling to leave these platforms.Source: §1 · pp. 1–2
2. Educational Artifact Development
The artifact uses a simplified, privacy-preserving LLM simulation and an interactive Sankey diagram to make the path from digital activity to recommendations traceable. Its pipeline ingests activity archives, generates and groups inferred interests, grounds recommendations in products, and visualizes the relationships.
2. Educational Artifact Development
- The simplified LLM-based simulation and interactive Sankey diagram are designed to help young adults trace digital activity through profiling to personalized recommendations.The artifact uses a locally hosted LLM and secure data handling so that data stays on an IRB-approved server and need not be manually inspected or shared with third parties.Source: §2 · p. 2
2.1. Step 1: Ingesting Users’ Online Activities
- Users anonymously upload at least four activity archives, which are normalized into records containing item, link, time, date, and action information.Sources can include browser, social media, location, search, news, shopping, and entertainment histories.Source: §2.1 · pp. 2–3
2.2. Step 2: Generating User Profiles
- A locally run open-source LLM processes sensitive behavioral traces to infer user interests and corresponding products, avoiding transmission to third-party LLM servers.The authors also cite reproducibility, auditability, predictable cost, and independence from API pricing and rate limits as reasons for local inference.Source: §2.2 · p. 3
2.2.1. Inferring Interests and Recommendations
- The model generates interests for individual and combined datasets, then merges similar inferences, clusters them into categories, and assigns sensitivity and commonness scores.The number of inferences per combination can range from 3 to 9; requests above nine produced duplicates. Users can filter by category and either score.Source: §2.2.1 · p. 3
2.2.2. Mapping LLM-generated Recommendations to Amazon Products
- A retrieval component matches each generic LLM recommendation to the top-1 semantically similar product in the Amazon Reviews 2023 dataset.Product titles and descriptions are embedded and indexed for similarity search before the match is attached as a grounded example.Source: §2.2.2 · pp. 3–4
2.3. Step 3: Producing the Sankey Diagram
- The three Sankey layers encode activity sources, inferred interests, and recommendations; links quantify their prevalence and filters let users inspect selected sources or inference attributes.For an inference to remain visible, all contributing sources must be enabled; recommendations linked only to hidden inferences are also hidden.Source: §2.3 · p. 4
3. Educational Artifact Evaluation using a Think-Aloud study
The authors evaluated the artifact in a pilot think-aloud study with six young adults. Participants examined the same visualization, based on the first author’s data, through tasks about comprehension, privacy, and usability.
3. Educational Artifact Evaluation using a Think-Aloud study
- Six young adults took part in one-hour think-aloud sessions conducted remotely via Zoom after IRB approval.Sessions were recorded and transcribed; participants were recruited by word of mouth.Source: §3 · p. 4
- All participants evaluated the same visualization, which was generated from the first author’s personal data rather than their own.This allowed the researchers to provide a consistent visualization without asking participants to upload personal data.Source: §3 · p. 5
- Seven tasks covered participants’ impressions and understanding, tracing sources and inferences with filters, privacy perceptions and actions, and suggestions for improving usability and comprehension.Participants explored data-source provenance and used sensitivity and uncommonness filters to examine how interactions contributed to inferences and recommendations.Source: §3 · p. 5
4. Think-Aloud Results
Participants said the Sankey simulation helped them understand how online activity relates to profiling and recommendations. They also identified limits in their understanding and in the visualization, and did not expect to substantially change their online behavior.
4. Think-Aloud Results
- Participants said the Sankey simulation helped them understand the links between online activities, user profiling, and personalized recommendations.They described the visualization as intuitive and said tracing from activities to recommendations, or backward to their sources, improved interpretability and conceptual trust.Source: §4 · p. 5
- Participants’ initial understanding was often incomplete or misguided, and the artifact helped them recognize that historical data can persist and inform recommendations.One participant connected old computer-building interests to later ads and recommendations after realizing that earlier activity remained relevant.Source: §4 · p. 5
- All participants noticed that distinct activities could jointly contribute to a seemingly unrelated inferred interest, though they initially could not explain how.When prompted, some loosely connected this process to cross-site tracking, but they did not articulate its mechanism.Source: §4 · pp. 5–6
- Participants said they might exercise more caution online but felt unable or unwilling to change their behavior meaningfully because digital ecosystems felt unavoidable.They described search, video, and navigation services as essential despite discomfort with behavioral tracking.Source: §4 · p. 6
- Participants proposed clearer labels and filters, explanations of sensitivity scores, and interaction cues; they also wanted context for links between seemingly unrelated activities and recommendations.Suggested changes included highlighting selected connections and distinguishing browser history from Google search history.Source: §4 · p. 6
5. Discussion
The discussion finds the artifact helps young adults understand how digital activity informs profiling and recommendations, while prompting privacy reflection without reliably changing behavior. Participants still struggled to understand cross-platform tracking, and the study notes limitations in the artifact and its evaluation.
5. Discussion
- The artifact helped young adults understand how digital activities jointly inform inferred interests and recommendations, and partly dispelled the misconception that tracking occurs only through hashtags.Participants nevertheless lacked a clear understanding of cross-platform tracking, even after substantial prompting.Source: §5 · p. 6
- Participants reflected on exercising more caution online, but most viewed tracked platforms as difficult to abandon, sometimes feeling helpless and potentially becoming defeatist about privacy practices.The authors suggest this may indicate a broader need to provide users with autonomy and agency over their data.Source: §5 · p. 6
- A larger sample could further solidify the artifact’s effectiveness, and the artifact prioritizes saliency over accuracy for young adults’ comprehension.Source: §5 · p. 6
- The simplified model could mislead in some instances, while using another person’s data likely set a lower bound on educational value; classroom scaffolding may sustain impact.Source: §5 · p. 7
6. Conclusion
The paper presents an educational artifact that simulates how digital activities shape online experiences. Its preliminary evaluation found that it promoted comprehension and privacy reflection, but users may not change behavior when they see no alternative to platform use or means of protecting their data.
6. Conclusion
- The illustrative model promoted comprehension and reflection about privacy implications, but users may not change behavior when they see no alternative to platform use or protecting their data.Source: §6 · p. 7
7. Declaration on Generative AI
The authors used Grammarly for grammar and spelling checks, then reviewed and edited the content and retained responsibility for the publication.
7. Declaration on Generative AI
- The authors used Grammarly for grammar and spelling checks, then reviewed and edited the content and retained responsibility for the publication.Source: §7 · p. 7