Data scientist
Mathematical and statistical foundations, Python, machine learning and its honest evaluation: the path from analytics to modelling.
Sign in to track how many of the 33 roadmap tasks you have done.
For students and analysts who want to go beyond describing data: to build models that predict and to know when to trust them. The roadmap goes through mathematical and statistical foundations, the Python ecosystem, machine learning and its evaluation, all the way to deployment and presenting results. By the end you will manage the whole cycle: from a question through data and a model to an honest evaluation and clear communication.
Steps
Step 1: Mathematical foundations
Linear algebra (vectors, matrices, matrix multiplication), the basics of calculus (derivative, gradient) and probability are the language in which all machine learning models are written. You do not have to prove theorems, but you must understand what happens when a model "trains". You are done when you can explain what a dot product, an eigenvector, a gradient and a conditional probability are, and why ML rests on them.
Topics on the portal
Resources
Step 2: Inferential statistics and experiments
Descriptive statistics says what is in the sample; inferential statistics says what follows from it for the whole population. Confidence intervals, hypothesis testing, p-values and experiment design are the foundation without which every "model improvement" may turn out to be noise. You are done when you design an experiment, evaluate it including a confidence interval, and know when a difference is both statistically and practically significant.
Topics on the portal
Tasks
Resources
Step 3: The Python ecosystem
NumPy, pandas, Matplotlib and Jupyter are the everyday tools of a data scientist; clean and reproducible code (virtual environments, Git, functions instead of copying cells) is what separates a notebook from a result you can trust. You are done when you can prepare a dataset with pandas, vectorise a calculation with NumPy instead of a loop, and hand over a notebook that a colleague can run from start to finish.
Tasks
Resources
Step 4: Machine learning basics
external resourcesSupervised learning (regression, classification), unsupervised learning (clustering, dimensionality reduction), and concepts such as training and test sets, overfitting and regularisation. More important than knowing fifty algorithms is understanding when a linear model is better than gradient boosting. You are done when you train and compare three different models on one dataset and explain why the winner won.
Resources
Step 5: Model evaluation
Accuracy is almost worthless on imbalanced data, so you need precision, recall, F1, ROC/AUC, the confusion matrix and, for regression, MAE or RMSE. On top of that come correct data splitting, cross-validation and the awareness that information leaking from the test set into training is the most common way of fooling yourself. You are done when, for a given problem, you choose a metric, defend it by the business impact of false positives and false negatives, and set up validation without data leakage.
Tasks
Resources
Step 6: Feature engineering
external resourcesThe quality of the input features decides the model more than the choice of algorithm. Scaling, encoding categories, handling missing values, features derived from dates and text, feature selection, and the discipline of doing it all in a pipeline so that the transformations on the test set do not differ from training. You are done when you build a scikit-learn Pipeline with preprocessing and a model and show how much better features improved the result.
Resources
Step 7: Deep learning basics
external resourcesNeural networks, backpropagation, optimisers, and why they now sit behind images, speech and language models. For most tabular problems you do not need them, but you must know when you do, and how to use a pretrained model instead of training from scratch. You are done when you train a small classification network in PyTorch, explain what happens in a single training step, and apply a pretrained model to your own data.
Resources
Step 8: MLOps and deployment
optionalexternal resourcesA model in a notebook earns nothing. Experiment tracking, versioning of models and data, serving a model through an API, drift monitoring and automatic retraining are what turn a prototype into a product. You are done when you package a model into a service with an API, its predictions are logged, and you can say how you will recognise that it needs retraining.
Resources
Step 9: Communicating results
The result of a model has to be understood by a manager who does not know the term AUC. Explainability (which features drive the decision), honest reporting of uncertainty and visualisations that do not mislead are as important as accuracy itself. You are done when you present the results of a project on three slides (problem, solution, impact) and answer the question "why should we trust it".
Topics on the portal
Resources