The GOAT Debate — YouTube Comment Network Analysis

Six stages as six separate notebooks, each writing its outputs to disk and each runnable on its own, so a change to the topic model does not mean re-scraping the data.

Sentiment scored twice — with VADER and with RoBERTa — so a lexicon method and a transformer can be compared on identical comments rather than one being trusted by default.

Topics extracted with both LDA and BERTopic, for the same reason: two models that disagree about what an argument is about is information, and a single model’s topic list looks authoritative whether or not it deserves to.

The comment graph analysed for betweenness centrality to find the accounts bridging the two camps, Louvain communities to find the camps in the first place, and assortativity to measure how much each side only ever replies to itself.

Influence modelled as cascades over the network, with a sensitivity sweep rather than one run — a cascade result from a single parameter setting says more about the parameter than about the network.

An earlier solo assignment in the same unit built the collection half of it: YouTube Data API scripts targeting the 2026 Australian Grand Prix across six query sets, and a text pipeline where the interesting problem turned out to be stopwords — the standard NLTK list left “f1”, “race” and “car” dominating every frequency count, so the domain-specific list had to be built by hand before the topics underneath were visible.