Machine learning postgraduate personal statement example

Personal statement example

I have spent six years watching machine learning arrive in my workplace from the outside. As a data analyst for a bus operator, my job is to explain why passenger numbers on a route fell in February, or how many vehicles a new school timetable will need. The methods I use are honest and limited: SQL aggregations, seasonal averages, occasionally a regression. When a supplier demonstrated a demand-forecasting product to us, I could follow the outputs but not the reasoning, and I disliked that. I could not tell whether the model was learning something real about travel behaviour or memorising the previous year's data. That gap between using a tool and understanding it is the reason I am applying for postgraduate study in machine learning.

My first degree was in mathematics, which I finished in 2016. I chose modules in probability, statistics and numerical methods, and my final-year project applied Markov chains to a simple queueing problem, including a discrete-event simulation I wrote in MATLAB. That project taught me that the interesting work usually sits in the assumptions rather than the algebra: my results were only as good as the memorylessness I had assumed about arrivals. Six years on, I recognise the same issue in the independence and stationarity assumptions behind the forecasting I do at work.

Because I have been away from formal study, I have prepared deliberately rather than relying on what I remember. Over the past eighteen months I have worked through a structured online course on machine learning, completing the assessed programming work in Python, and I have re-read my undergraduate linear algebra notes alongside Strang's lectures until singular value decomposition felt like something I could use rather than recite. I have also been reading An Introduction to Statistical Learning, which suits me because it is careful about bias and variance before it is enthusiastic about methods. Working through the cross-validation chapters changed how I evaluate my own reports at work; I now split historical data by time period rather than at random when I test a forecast, which had been quietly flattering my results.

My own projects have been small and grounded in what I know. Using published timetable and journey data, I built gradient-boosted and linear models to predict arrival delay at a set of stops, and found that the tree-based model gained most of its advantage from features I had engineered by hand from time of day and school-term calendars rather than from the algorithm itself. I have also experimented with a simple neural network for the same task using PyTorch, mainly to learn the mechanics of training loops, and found it harder to justify than the simpler model. Reaching that conclusion, rather than assuming the more complex method must win, feels like progress.

Work has given me habits I expect to be useful. I am used to versioning code so that a colleague can reproduce a figure months later, to documenting where a dataset came from, and to explaining an uncertain result to operations managers who need to make a decision anyway. Coaching a junior netball team on Saturdays has made me better at breaking an idea into steps and noticing when someone has stopped following.

I am particularly interested in probabilistic modelling and in the evaluation of models deployed on data that shifts over time, since public transport demand plainly does. Longer term I would like to work on forecasting and optimisation in the transport or public sector, where I understand the operational constraints and where interpretable, well-tested models matter. Returning to full-time study is a considered decision, and I am ready for the mathematical demands of it.