Published Updated
Tags research machine-learning oss
What I Learned From an Earlier BiLSTM Thesis Direction
A retrospective on an earlier OSS-abandonment thesis direction, what the BiLSTM framing taught me, and why I changed the research question when the literature changed the evidence.
Update — September 2026: this post documents an earlier thesis direction. My current research focuses on vulnerable-dependency remediation viability in OSS ecosystems. I’m keeping these notes because the change itself is part of what I learned: a research plan should change when stronger evidence changes the question.
What I was exploring
My first thesis direction started from repository activity and asked whether sequential signals could help identify OSS abandonment risk. BiLSTM was the candidate model because commit, issue and contributor activity are temporal rather than isolated observations.
That framing helped me learn how quickly a model choice can shape the rest of a study: dataset construction, labels, observation windows, baselines and evaluation all begin to orbit the model if I am not careful.
Why I changed direction
The literature review made the original problem less compelling as a thesis contribution than I had assumed. Instead of defending the first idea, I changed the research question and moved the model family back to an open decision.
The current work studies vulnerable-dependency remediation viability: whether historical upstream maintenance signals add useful prospective information about remediation timing beyond vulnerability, remediation and consumer-project characteristics.
What carried forward
The earlier direction was still useful. I kept several lessons:
- temporal data needs explicit observation windows and leakage controls;
- labels and outcome definitions deserve as much scrutiny as the model;
- stronger models are not automatically better evidence;
- baselines matter because the useful question is whether a new signal adds value;
- a negative or non-incremental result can be informative if the design is sound.
The part I want to keep visible
Changing the thesis direction was not lost work. It was a correction produced by reading more, comparing the problem against existing research and being willing to revise my own hypothesis. That is also how I want to work as an engineer: make assumptions explicit, test them, and change the plan when the evidence no longer supports them.
The research page now reflects the current thesis direction and the decisions that are still intentionally open.