Chat-Driven Games and Viewer Retention

ISEF Category: Technology Enhances the Arts

Subcategory: Games  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Most streams lose viewers in the first few minutes. A game that turns chat into storms, allies, or hazards can change that fast. The trick is not just reacting to chat, but deciding when the game should react. Your project asks whether that extra feedback loop keeps people watching longer.

What Is It?

This project studies a game that listens to viewer chat and turns those messages into gameplay events. A sentiment model guesses whether a message feels positive, negative, or neutral. A topic model groups messages by what they are about, such as strategy, jokes, or complaints. Then the game maps those signals to events like power-ups, storms, or enemy spawns.

Think of it like a DJ set for gameplay. If you change the song too often, the crowd gets lost. If you change it too little, the room gets bored. Your job is to test the balance. Token buckets help with that balance. A token bucket is a simple rate limit. It stores a limited number of tokens, and each chat-triggered event spends one. That keeps chat from overwhelming the game while still making the stream feel live.

The research side is viewer retention. That means how long people stay, how often they return, and when they leave. You can compare the chat-reactive version with a static stream, then see whether different pacing rules change watch time, chat activity, or viewer drop-off.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with real users, clear metrics, and a controllable system. You are not guessing whether chat interaction helps. You can measure retention, event frequency, and response patterns. The topic also connects to real problems in streaming, game design, and human-computer interaction. A student can learn text analysis, experimental design, and basic statistics without needing a wet lab.

Research Questions

• How does chat-to-event rate limiting affect average viewer watch time?
• What is the effect of sentiment-triggered events on viewer return rate?
• Does topic-based event mapping keep viewers active in chat longer than random event mapping?
• To what extent does event frequency change viewer drop-off during the first 10 minutes of a stream?
• Which event type, ally, hazard, or storm, leads to the highest chat engagement?
• How does combining sentiment and topic signals compare with using sentiment alone for viewer retention?

Basic Materials

• Laptop or desktop computer with a microphone and webcam
• Free streaming software such as OBS Studio
• Game engine such as Unity or Godot
• Python installed on the same computer
• Text chat log source from a private test stream or scripted viewers
• Spreadsheet software such as Google Sheets or Excel
• Timer or screen recording tool
• Consent form template for any human viewers

Advanced Materials

• Access to a small server or cloud machine for live chat processing
• Python libraries for NLP such as spaCy or scikit-learn
• A sentiment model or API with a free tier
• A topic modeling library such as BERTopic or scikit-learn LDA tools
• Database such as SQLite or PostgreSQL for event logs
• Statistical software such as R or Python with pandas, SciPy, and statsmodels
• Version control with Git
• Optional eye-tracking or engagement logging tools if your school or lab has them

Software & Tools

• OBS Studio: Records or streams each condition so you can compare viewer behavior across runs.
• Python: Runs the chat parser, event logic, and data cleaning scripts.
• Google Sheets: Tracks retention, event counts, and simple summary statistics.
• R: Tests whether changes in retention are statistically meaningful.
• ImageJ: Can help if you need to measure on-screen visual changes in recorded stream footage.

Experiment Steps

1. Define the viewer behavior you want to measure, such as watch time, chat rate, or return visits.
2. Choose one chat signal to test first, then decide how that signal will trigger an in-game event.
3. Set up a control version with no chat-reactive events so you have a fair comparison.
4. Design a token bucket rule that limits how often chat can change the game.
5. Plan how you will log each event, each chat burst, and each viewer outcome.
6. Pick the statistics you will use before you run the experiment, so you do not chase noise later.

Common Pitfalls

• Treating every chat message as equal, which makes spam drown out meaningful signals.
• Using sentiment scores without checking whether sarcasm or slang breaks the model.
• Changing event frequency and event type at the same time, which hides the real cause of viewer retention changes.
• Measuring retention with only one stream session, which gives you too little data to trust the result.
• Forgetting to log the exact moment each event happens, which makes it hard to connect viewer behavior to game changes.

What Makes This Competitive

A stronger version of this project would separate design choices cleanly. You could compare sentiment-only, topic-only, and combined systems, then test more than one token bucket rule. That gives you a real algorithm comparison, not just a demo. You can also make the project stronger by using preplanned metrics, careful retention analysis, and a clear control stream.

Project Variations

• Test the same chat-reactive system on a rhythm game instead of a strategy game to see whether genre changes retention.
• Replace live viewer chat with a scripted message set so you can isolate how the model, not the audience, affects event pacing.
• Compare English-only chat processing with multilingual chat processing to see whether language mix changes classification quality and retention.

Learn More

• MIT OpenCourseWare, Introduction to Computer Science and Programming in Python: Search MIT OpenCourseWare for Python programming basics and small data projects.
• Google Scholar: Search for papers on livestream engagement, chat interaction, and human-computer interaction.
• PubMed: Search for review articles on attention, feedback loops, and digital engagement if you want a psychology angle.
• NIH PubMed Central: Find free full-text research articles on user behavior, sentiment analysis, and interactive media.
• Python documentation: Read the official docs for libraries such as pandas, scikit-learn, and asyncio.
• OBS Studio help center: Learn how to record or stream each condition in a consistent way.