National University of Piura · Bachelor Thesis

Predicting the Abandonment State of OSS Repositories using BiLSTM Neural Networks

In Progress

▸ Problem statement

A significant number of Open Source Software (OSS) repositories become abandoned over time, leaving dependent projects exposed to security vulnerabilities, missing updates, and technical debt. Early detection of abandonment risk is a critical yet underexplored challenge in software ecosystem health.

▸ Research hypothesis

“Can Bidirectional LSTM (BiLSTM) networks, trained on historical activity metrics from GitHub repositories (commits, issues, pull requests, contributor patterns), predict the abandonment state of an OSS repository with higher accuracy than traditional ML baselines?”

▸ Methodology (MSR)

Mining Software Repositories (MSR) — extracting temporal sequences of activity signals from the GitHub API (GH Archive), engineering features that capture contribution decay, issue resolution rates, and bus factor trends, then training BiLSTM models to classify future abandonment probability.

Data pipeline

GitHub API / GH Archive Raw Repo Activity Data Feature Engineering Temporal Sequences BiLSTM Training Abandonment Classifier

▸ Engineered features

F01 Commit frequency (weekly rolling avg)

F02 Unique contributors per time window

F03 Open issue aging (days unresolved)

F04 PR merge rate & close rate

F05 Days since last commit

F06 Bus factor estimate

F07 Issue response time

F08 Release cadence

▸ Dataset

Curated dataset of GitHub OSS repositories spanning multiple domains, labeled by abandonment state using historical activity thresholds (commit frequency, last contribution date, open issue aging).

▸ Model architecture

Input Time-series of repo activity signals (T steps)

Layer 1 BiLSTM — 128 units forward + backward

Layer 2 BiLSTM — 64 units with dropout (0.3)

Layer 3 Dense — 32 units, ReLU activation

Output Sigmoid — P(abandoned) ∈ [0,1]

Loss Binary Cross-Entropy + class weights

Optimizer Adam (lr=1e-3)

▸ Evaluation metrics

ACC Classification accuracy vs. Logistic Regression & Random Forest baselines

F1 F1-Score on imbalanced abandonment classes

AUC AUC-ROC for abandonment probability ranking

LOSS Binary Cross-Entropy loss curves for BiLSTM training convergence

▸ Current status

In Progress

Actively training BiLSTM models and comparing against ML baselines. Results will be published upon thesis completion.

Keywords

BiLSTM OSS Abandonment MSR Deep Learning GitHub Mining Repository Health Time Series Classification

Interested? Reach out

Collaborations, questions or opportunities.

Contact