Capstone: Build the Features

A model learns from a table with one row per thing it predicts, here one row per customer. The columns of that table are the features. customers.csv already has one row per customer, but orders.csv has about 1,500 order lines. The starting code cleans the customers, gives each order line a value (quantity times unit price) and sums the lines of each order into by_order. You boil those orders down to two new features per customer: the average order value and the share of cancelled orders.

From orders to customers: comparing a column with a value, like orders["status"] == "paid", gives back one True or False per row, and the mean of True/False is the share of True. groupby("customer_id")[[...]].mean() gives back one row per customer with the mean of each column you name. Setting .columns to a list renames the columns, in order, and merge(other, on="customer_id") gives back a table where each row also gets the columns of the matching row in other. For example, per category:

by_category = orders.groupby("category")[["quantity", "unit_price"]].mean()
by_category.columns = ["avg_quantity", "avg_price"]
orders = orders.merge(by_category, on="category")

X, y and the split: drop(columns=[...]) gives back the table without the columns you list: X keeps the features, and y is the label, bought_again. train_test_split(X, y, ...) from scikit-learn gives back four parts, X_train, X_test, y_train, y_test. test_size=0.25 keeps a quarter of the rows for the final test, and stratify=y keeps the same share of 1s in both parts.

When you're done, the check after your code prints:

     city   age newsletter  visits  orders  spent  days_since_last  returned_item  avg_order_value  cancelled_share
0    Rome  52.0        yes      11       6    720               21              0          120.000              0.0
1  Madrid  41.0        yes       7       3    201               12              0           67.000              0.0
2  Madrid  51.0         no       5       3    257                1              0           85.667              0.0
Average order value: 70.24 cancelled share: 0.074
Rows: 300 train: 225 test: 75
Bought again in train: 0.596 in test: 0.6

Your task: add a True/False "cancelled" column to by_order, build per_customer with each customer's mean "value" and "cancelled", rename its columns to "avg_order_value" and "cancelled_share", and merge it into customers. Then make X (without "customer_id", "joined" and "bought_again"), y, and a stratified split with test_size=0.25 and random_state=0.

Assignment
Follow the Coding Tutorial to add the two order features to customers and split it into training and test rows.

Hint
Group by_order, not orders: it has one row per order, so an order with three lines counts once. Never put the label bought_again in X.