Problem
Understanding how YouTube's recommendation system shapes viewers' ideological exposure over time requires modeling recommendations as a network and simulating how a viewer moves through it — not just looking at individual video metadata.
Approach
I built an end-to-end Python analytics pipeline that analyzes 7,079 YouTube channels and 401,384 recommendation relationships between them, constructing a weighted network for large-scale ideological drift analysis. On top of that network, I developed reproducible Monte Carlo-style weighted random walk simulations to model recommendation behavior, measuring changes in ideological direction, extremity, assortativity, and clustering as a simulated viewer follows recommendations. The codebase is test-driven, with 72 automated tests backing the simulation and analysis code.
Key decision
The full technical-decision write-up — the real alternatives considered for network construction and simulation design, and why they were chosen — is pending until the repository is cleaned up enough to reference publicly. This section will be filled in once that pass is done.
Outcome
The pipeline produces reproducible simulation results across the full 7,079-channel network, with documentation and a 72-test suite backing the analysis. A public write-up of the specific findings is still pending.
What I'd do differently
Pending the same repo cleanup as the Key decision section above — a retrospective written against the actual code and history rather than from memory.
This write-up is a placeholder. It's blocked on reorganizing the GitHub repo into a presentable, linkable state; the repo isn't public yet and the full findings write-up isn't finished. The structure above will be filled in once those are ready.