Lesson 2 of 11

Data: features, target and splitting

Every model starts with a table. Each row is an example (a customer, an invoice, a month), and the columns come in two kinds: features are what we know, and the target (or label) is what we want to predict.

Store	Orders / month	Weeks since last order	Region	Left? (target)
Al Waha	8	1	Riyadh	No
Al Salam	2	7	Jeddah	Yes
Al Noor	5	4	Dammam	No

Why split the data? Testing a model on the examples it learned from is like a student who memorised the exam with its answers: full marks, no understanding. So we hold back part of the data (usually 20%) as the test set, learn from the rest (the training set), and then see how the model does on examples it has never seen.

PYTHON

months = ["Jan", "Feb", "Mar", "Apr", "May", "Jun", "Jul", "Aug", "Sep", "Oct"]
cartons = [410, 395, 430, 450, 470, 520, 610, 590, 480, 505]

cut = int(len(cartons) * 0.8)   # 80% for training
train = cartons[:cut]
test = cartons[cut:]
print("Train:", len(train), "months")
print("Test: ", len(test), "months", test)

Watch outWith time-based data (monthly sales, say) keep the latest period for testing, as above, because in real life you predict the future from the past. If rows are independent (customers, say), shuffle them randomly before splitting.

Data matters more than the algorithm. A model learns from what you give it: if the sales data has duplicates, mistyped prices or missing months, it learns the mistakes. The classic rule: garbage in, garbage out. That's why most of an ML project's time goes into preparing and cleaning data, and where your SQL skills pay off.

Data: features, target and splitting

Exercises