The Value of GitHub Stars and Detecting Fake Stars Using Unsupervised Clustering

GitHub stars are often treated as a vanity metric, yet they influence technical selection, investment decisions, and recruitment, making them a target for manipulation.

The market for buying stars is real: services such as GitHub24 and Baddhi Shop sell thousands of stars for as little as $64, while higher‑quality active accounts can cost up to $85 for a few hundred stars.

Simple heuristics can identify low‑quality fake accounts: followers < 1, following < 1, few public repositories, and empty profile information. GitHub periodically removes such accounts.

Several tools—astronomer, fake‑star‑detector—automate this basic detection.

For more sophisticated fraud, Dagster’s open‑source platform applies unsupervised clustering on behavioral features (code commits, PRs, starring patterns, profile edits) to represent each GitHub account as a high‑dimensional vector. Accounts that cluster together with known fake accounts are flagged as suspicious.

Experiments on a repository where all stars were purchased show clear separation: genuine users form a blue cloud, confirmed fake accounts appear in red, and the clustering algorithm highlights additional yellow accounts as likely fakes.

Applying the same analysis to a legitimate project (Dagster) reveals minimal overlap with fake accounts, while a cryptocurrency project (okcash) shows 97 % of its stargazers as suspected fakes, potentially harming market confidence.

Readers interested in reproducing the detection pipeline can follow the Dagster tutorial linked in the article.