Machine Learning With Python: From Zero to a Working Model

Every lesson uses the same few words: feature, label, model, training and predicting. If those words are fuzzy, the code will feel like magic. This course ends with a model that predicts home sale prices from a real table, plus a plain-English report about it. Each module ends with a small project that adds one piece to that final work. Each lesson has a short task and a three-question quiz, so read, do the task, then take the quiz. Here you learn the words with a pencil and a tiny table, before you type any code.

Normal software follows rules that a person wrote. A price calculator might say, 'add 50 dollars for each bedroom.' Someone had to guess that rule.

Machine learning flips this around. You give the computer many past examples, each with the right answer. The computer looks for a pattern in them and builds its own rule. That rule is called a model.

These words are used all course. Each is defined once, here.

• Row: one example. In a table of home sales, one row is one home.
• Feature: one fact you know before you predict. The size of a home is a feature.
• Label (also called the target): the answer you want to predict. The sale price is a label.
• Training (also called fitting): the learning step. The model studies rows where the label is already known.
• Predicting: the using step. The model receives the features of a new row and returns a guess for the label.

There are two common kinds of question.

• Predict a number, such as a price or a delivery time. This is called regression.
• Predict a category, such as yes or no. This is called classification.

Use these steps to frame any problem.

1. Write the question as one sentence: 'Given these facts, predict this answer.'
2. List the facts you will know on the day you need the prediction. These are your features.
3. Check that you have past rows where the answer is known. That column is your label.
4. Decide if the answer is a number or a category.

One warning to keep in mind. A model only knows what its examples show. It finds patterns. It does not know why they exist, and it cannot see the future.

Example (illustrative numbers). A small table of past home sales:

The question is: 'Given the size of a home, predict its price.' Size is the feature. Price is the label. The answer is a number, so this is regression.

Now predict the price of a new home of 1,450 sq ft. We use a simple pattern-finding method.

1. Work out how far each past home is from 1,450 sq ft. The 1,500 home is 50 away. The 1,300 home is 150 away. The 1,700 home is 250 away. The rest are farther.
2. Keep the two closest homes: 1,500 (price 190) and 1,300 (price 170).
3. Average their prices: (190 + 170) / 2 = 180.
4. The prediction is 180 thousand dollars.

That is a real model. Its training step is storing the table. Its predicting step is the four moves above. A computer does the same job on thousands of rows and many features at once.

• Putting the answer inside the features. It happens because the price column sits in the same table as the size column. Fix: name the label first, then build the feature list without it.
• Expecting the model to explain why. A model finds patterns. It does not know reasons. Fix: treat every prediction as an estimate, and later lessons show how to check the estimates.
• Using a fact you will not have in real life. For example, a final invoice total cannot help predict a complaint that happens before the invoice exists. Fix: for each feature, ask, 'Will I have this on the day I need the prediction?'

You can take any table, write the question as 'Given ___, predict ___', circle the features, underline the label, and say whether the question is regression or classification.