GamePulse

Brief

Without access to a game’s internal data, public reviews are the closest you get to the players’ own voice. GamePulse collects reviews from Steam, cleans them, and analyses sentiment, topics and kinds of player, with the results laid out in a front-end workbench.

Process

The first version seemed to have everything, but many of its conclusions had never been checked. The second round did one thing: give every result a reason to be believed.

The topic model was the first to look wrong. The corpus turned out to mix more than twenty languages, and LDA was really grouping by language: the normalised mutual information between topic and language was 0.42, and four of ten topics were simply the Russian, Spanish or Portuguese reviews. Modelling each Steam language tag separately made those clusters disappear.

One result went against intuition. After filtering, NPMI topic coherence fell from 0.26 to 0.11, because NPMI rewards words of one language clustering together. So it can’t be read alone, only next to the mutual information between topic and language.

The sentiment model was tested against Steam’s own “recommended” flag as a weak label: 70.3% agreement on English reviews, falling to 32–44% for Spanish, Portuguese and Russian. That gap of twenty to thirty points is the price of running an English-only model on a multilingual corpus.

Outcome

The workbench can speak to reputation, pain points and kinds of reviewer. Retention, spending and ROI are beyond what public reviews can show, so it doesn’t try. With fewer than 300 reviews it gives a preview, not a full conclusion. The code is on GitHub.