My PhD Thesis Outline - Planning for My Final Year as a Computer Science PhD Student

With just 15 months left of funding, a computer science PhD student in machine learning lays out a tightly sequenced plan to finish research and shift into thesis writing—aiming to “wrap up” core research within about six months so roughly nine months can go to drafting, revising, and finalizing the thesis.

The thesis centers on marathon training analytics using Strava data. The work focuses on turning complex, messy time-series training logs—each run session broken into 100-meter intervals with distance, time, pace, elevation, and heart-rate—into models that can (1) predict race performance, (2) quantify how training disruptions affect outcomes, and (3) recommend training plans to improve performance. The scale of the dataset grows dramatically across the PhD: early experiments used about 20,000 runners over 16 weeks, while later work expands to a couple million runners and tens of millions of sessions. That expansion forces new data engineering skills: extracting, cleaning, and re-running analyses at a much larger scale.

Early progress came quickly. In the first year, the student built features from training history to predict race times and produced two first-author conference publications, including work on training recommendations (what a “next week” might look like depending on whether a runner wants to improve or ease off). After the pandemic, the research focus shifted toward training disruptions—examining how breaks in training influence marathon performance—supported by multiple mini-projects and later journal submission.

Structurally, the thesis is planned around the experiments rather than a purely linear “methods/results” flow. The outline includes an introduction chapter, a literature review, and a dedicated dataset chapter that documents how raw Strava data becomes usable modeling inputs, including the cuts and cleaning decisions required for a messy dataset. The core contribution is then split into three main chapters: (1) feature engineering and performance prediction, (2) training disruptions using more data-science-heavy analysis with some machine learning, and (3) training recommendations framed like a recommender system for marathon running.

A key practical challenge is validation for recommendations. Live user studies are unlikely, so validation is expected to rely on generating sample training plans and recruiting people to judge whether the plans are sensible and understandable—more “can humans follow this?” than “does it improve outcomes in the real world?”

For the remaining year, the student’s calendar is designed to front-load research during a less teaching-heavy summer. June priorities include finishing “camera-ready” material for an accepted conference paper and completing analysis for an additional training-disruptions piece that could become its own publication. The next phase targets feature engineering and representation learning for performance prediction, followed by a user-evaluation step to test explainability and usability of training plans. The final recommender-system work is scheduled for roughly October through December. Throughout, the student plans to write iteratively—capturing notes and drafts early—while avoiding premature full write-ups that might require rerunning models if the dataset changes again.