StimScapes
Computational Media Research @ UCSC
My particular project was done under Professor Kate Ringland, an Assistant Professor at the University of California, Santa Cruz. My PhD mentor was Yihe Wang. I worked in a group of three high school interns.
Exploring Neurodivergent-Led Media Practices to Reframe Design Principles for Neurodiverse Multimodal Learning
We aimed to understand the neurodivergent community's online viewing habits to learn the design principles behind creating stimming experiences. The final goal of the research was to design a game that used these design principles to promote stimming.
Motivation
Autism: A neurological development condition affecting people on a wide spectrum.
Stimming: Self-stimulation through repetitive/unusual movements, often seen in individuals on the autism spectrum. It serves as a way to calm emotions or for other purposes and is sometimes viewed negatively in society.
Background
Method Overview
How can computational technology support stimming experiences for individuals with autism?
What are the predominant stimming behaviors exhibited by individuals with autism, and how do they contribute to emotional regulation and sensory processing?
What existing computational technologies are currently being employed to assist individuals with autism in their stimming practices?
What design principles should be considered when developing computational tools or applications to facilitate and enrich stimming for individuals with autism?
What are the cultural and societal factors that may influence the acceptance and adoption of computational technology as a means of supporting stimming for individuals with autism?
Research Questions
Hashtag Network
GOAL:
Extract and cluster hashtags to understand thematic principles in stimming behavior.
APPROACH:
Scrape TikTok for data (73 creators, 20,523 videos).
Extract, parse, and cluster 505 hashtags.
Give co-occurrence score (times 2 hashtags appear in the same video)
Use BERT Model for semantic scoring
Optimize clusters and analyze results.
Similarity Score Metrics
CO-OCCURENCE
Given hashtags i & j —> co(ij) = time i & j co-occur
Sim(ij) = [Co(ij)/Deg(i)] + [Co(ij)/Deg(j)]SEMANTIC
[1] Segment Hashtags
[2] Using BERT Model get word embeddings
[3] Cosine Similarity
Co-occurence network
Optimization
An optimal cluster will have a higher similarity between any two pairs within the same cluster.
Least Similarity - Average of the least sim(ij) of each cluster
Average Similarity - Average of the all sim(ij) within each cluster
Reporting Accuracies + Comparisons
Semantic outperformed
Cluster Labeling
In-Depth Video Analysis - Dovetail
[1] Qualitative Analysis
In each video, we briefly explained stims then highlighted based on a color code.
Here you can see the stims that were exhibited in a singular video
[2] Categorical Analysis
We separated descriptions into 3 groups (types of stimming, video content, portrayal methods)
By doing this we were able to fully describe each video and see similarities between videos.
[3] Statistical Analysis
Within each of the colors, there are specific descriptions.
When a phrase is highlighted it is automatically linked to the color + description, allowing us to keep track of the video counts.
Findings
Types of Stimming - Common: vestibular/proprioceptive and tactile stimming.
Video Content - Common: types of stimming, informational content, and stimming tools/tips.
Portrayal Methods - Common: video/demonstrations and music; provide a multi-sensory experience.
Prototyping
Figma wireframe
Our data analyses revealed a crucial need to focus on:
[1] Promotion of self-acceptance,
[2] Support for diverse preferences, and
[3] Educational/informative content
Using our newly acquired design principles, we prototyped a mobile application called StimScapes, an educational game designed to promote stimming amongst neurodivergent children.