Data Science & Machine Learning
Data science is the work of using data to help generate insights that help make good decisions. Machine learning is a subset where you train a model on historical data so it can make predictions about new data. Both require looking back at information from the past to help inform the future. These steps cover both, with labels showing which are most relevant depending on where you are in your career.
Why Python specifically
Python is the dominant language used in data science because of its ecosystem. It’s a flexible, all-purpose language that’s capable of high-level work, while also having a low barrier to entry. It’s well-supported for machine learning, as well as data analysis and visualisation.
What else should I learn?
There are three Python-specific tools that are used in varying levels in most data roles: Pandas, NumPy and Jupyter. Pandas handles tabular data (like a programmable spreadsheet), while NumPy handles numerical operations on arrays efficiently. Jupyter notebooks let you write code, see results, and annotate your thinking in the same document. Learn to use these and you’ll find you can manage most data work.
What to practise and how to practice it
Exploratory data analysis is the first thing you’ll do with any new data set in any data role. Start practicing by getting a real dataset (Kaggle has hundreds of free ones), then load it into a Pandas DataFrame and start exploring it. Count rows, check for missing values, filter by a column, group and aggregate.
What you actually need to know
While there’s more to data science than maths, and you certainly don’t need a maths degree to be a good data scientist, there’s no getting away from the fact that it’s important to really understand a few core mathematical principles. Get to know the difference between mean and median, but also what the difference tells you about your dataset. Variance and standard deviation will help you understand the range that your data is spread among. Learn distributions, what a normal distribution is and why it comes up everywhere. Make sure you understand the difference between correlation and causation, and how to interpret data so these two things don’t get confused.
How things can go wrong
Many of the most common mistakes in data work come from human error in misunderstanding and misinterpreting statistics. This could come in the form of A/B test results declared significant when they are not, or an average used to describe a highly skewed dataset. These are not edge cases: they happen regularly in real jobs. Having a good grounding in statistics methodology and interpretation is the foundation that lets you catch and avoid mistakes like these.
Why SQL before machine learning
Most real-world data is stored in databases. Before you can model it, you need to retrieve it, filter it, aggregate it, and join it with other tables. SQL (Structured Query Language) is the standard programming language that does all of this. Since data analysis requires interacting with databases on a daily basis, SQL is frequently the single most-used skill in a data analyst or data scientist role.
What to focus on
Master SELECT, WHERE, GROUP BY, and JOINs first. Then learn window functions (RANK, ROW_NUMBER, LAG, LEAD), which are used to calculate things like running totals, rankings, and period-over-period comparisons. These come up in almost every take-home technical task for data roles.
What is supervised learning?
Supervised and unsupervised learning are the two primary pillars of machine learning, though supervised learning is the most common type in commercial settings. Supervised learning involves having a dataset where you know the answer (also known as the label), and you train a model on historical examples so it can predict the answer for new examples it has not seen. Classification involves predicting discrete labels, while regression focuses on estimating continuous outcomes. In practice, classification could look like the information that allows your email provider to determine what is spam, while regression is the data that allows predictions to be made about house prices based on property size and location.
The evaluation problem
A model that memorises the training data perfectly, learning its noise and fluctuations, is not useful because it will not know how to deal with the randomness of new data, and the model will fail. This is called overfitting. Learning how to split data into training and test sets, and how to measure model performance properly (accuracy, precision, recall, AUC), is as important as knowing how the model itself works.
The core idea
Loosely modelled on the neurons in the brain, a neural network is a function that transforms inputs into outputs through layers of learned parameters called ‘weights’. During training, the network makes predictions, compares them to the correct answers, measures the error, and adjusts the weights to reduce that error. This adjustment process (known as backpropagation and gradient descent) is repeated millions of times until the network performs well.
What to actually build
Transformers or large language models are not a good place to start. Instead start by training a small network to classify handwritten digits using the dataset from the Modified National Institute of Standards and Technology. It is small enough to run on a laptop and it teaches you the full training loop: data loading, model definition, loss function, optimiser, and evaluation, all of which are essential to understanding the core principles of neural networks.
Why deployment is a separate skill
Most data science courses stop at training the model, but most data science jobs require you to know what happens after. Training a model that performs well in a notebook is not the same as running that model in a production system. The model needs to receive new data, return predictions quickly, handle errors, and be updated when performance degrades. Deployment and observing the model in practice is a key skill, and one that should not be overlooked.
What to learn
Learn how to save and load a trained model. Learn how to wrap it in a simple API (FastAPI is a good choice) so other systems can call it. Deploy it somewhere (Hugging Face Spaces and Render both have free tiers). Understand what model monitoring means: tracking prediction quality over time to catch when the model becomes stale.
The skill most data roles under-hire for
Data science has a communication problem. Although technically strong analysts produce correct findings, these findings may ultimately end up changing nothing because the analysts cannot explain them clearly, or communicate their significance. The work is not done when the model is trained or the analysis is complete. It is done when the person making the decision understands what you found, why they should believe it, and what it means for their decision-making.
What this looks like in practice
See communication as a core part of the job. This means learning to make charts that show one thing clearly rather than everything at once, writing the conclusion at the top, not the bottom, and using plain language for technical concepts. Practice explaining your work to someone who does not know what a p-value is. The truth is that most data jobs involve a lot of writing, but most courses do not prepare you for it.
