Predicting OSS Abandonment With BiLSTM: Notes From My Thesis

The problem

Open-source repositories die quietly. A project that looked healthy a year ago can end up with no maintainer, no merged PRs, and a growing pile of unaddressed security issues — and by the time that’s obvious, dependents have already built on top of it. My thesis asks a narrower question: can we predict abandonment risk early, directly from a repository’s activity time series?

Why BiLSTM

Commit frequency, issue resolution rate, and contributor churn are all sequential signals — what happened in month t depends on what happened in months before it, and abandonment rarely happens in a single step. A Bidirectional LSTM reads that sequence in both directions, which matters here: early warning signs (a maintainer going quiet, PRs piling up unreviewed) are easier to weigh correctly when the model has context from both before and after a given point in training data.

Data & features

Using GH Archive as the source, I extract per-repository monthly time series:

• Commit count and unique committers
• Issue open/close rates
• PR merge latency
• A rough “bus factor” proxy from contributor concentration

Repositories are labeled abandoned/active using activity-threshold heuristics on last-contribution date and open-issue aging — the same kind of signal maintainers already use informally, just made explicit.

Baseline comparison

The BiLSTM is benchmarked against Logistic Regression and Random Forest on the same features, evaluated with F1 on the (imbalanced) abandonment class and AUC-ROC for ranking risk. The goal isn’t just “does deep learning win” — it’s whether the sequence structure the BiLSTM can exploit actually adds signal over models that only see aggregated snapshots.

Full methodology and current results are on the research page — this post will get a follow-up once training converges on the full dataset.