Module 1: ML Foundations

What Machine Learning Is and Is Not

Machine learning is everywhere in business conversations, which makes it easy to overestimate or misunderstand. At its core, machine learning is a way of building software that learns patterns from examples instead of following rules a programmer wrote by hand. Knowing precisely what it can and cannot do will help you choose good projects, set realistic expectations and avoid expensive failures.

Rules versus learning from examples

Imagine a lender in Nairobi deciding which small business loan applications to approve. A traditional program would encode explicit rules: "approve if monthly revenue is above a threshold and there are no defaults in the past year". Someone has to invent, test and maintain every rule.

A machine learning approach instead takes thousands of past applications together with what actually happened (repaid or defaulted) and finds the patterns that best separate the two outcomes. The output is a model: a function that takes the facts about a new application and returns a prediction, such as a probability of repayment.

More formally, a model learns a mapping from features (the inputs, such as revenue, business age and sector) to a target (the outcome you want to predict). Learning means adjusting the model's internal parameters so its predictions on past examples are as accurate as possible, in a way that should carry over to new examples.

What machine learning is good at

• Prediction from many factors: estimating demand, delivery times, prices or the chance of an event.
• Classification: sorting emails, transactions, images or support tickets into categories.
• Finding structure: grouping similar customers or spotting unusual transactions.
• Ranking and recommendation: ordering products, search results or leads by likely relevance.

It is most valuable when the pattern is too complex to write as rules, when there are plenty of historical examples with known outcomes, and when the decision is repeated often enough that small improvements add up.

What machine learning is not

• Not magic or understanding. A model finds statistical patterns in its training data. It does not understand your business, and it can be confidently wrong on situations unlike its data.
• Not a fix for bad data. If past records are incomplete or biased, the model learns those flaws. "Garbage in, garbage out" applies fully.
• Not proof of cause and effect. A model may find that customers who use a certain feature churn less, but that does not show the feature causes loyalty.
• Not always necessary. If a clear rule or a simple average does the job, use it. Simpler solutions are cheaper to build, explain and maintain.
• Not "set and forget". The world changes. Prices, customer behaviour and regulations shift, and model accuracy can decay over time. This is often called drift.

Worked example: is this a good ML problem?

Consider three hypothetical proposals from a distribution company:

A useful checklist before any project: Is there a clear target you can measure? Do you have enough historical examples with that target recorded? Will a prediction change a decision? Can you measure whether the model helped?

The vocabulary you will use in this course

• Dataset: a table where each row is an example (a customer, a transaction) and each column is a feature or the target.
• Training: fitting the model to examples.
• Inference or prediction: using the trained model on new data.
• Generalisation: how well the model performs on data it has never seen, which is what really matters.

Common mistakes to avoid

• Starting with an algorithm instead of a business question and a measurable target.
• Ignoring a simple baseline, such as last month's average, which the model must beat to be worth using.
• Treating predictions as facts rather than estimates with uncertainty.
• Forgetting the people affected, especially in lending, hiring or pricing, where errors and bias have real consequences.