Skip to content

Exercises — Week 1 — NumPy: Fast Math on Whole Columns

What you are building

Feature totals as a 1-D array, a user-normalized usage matrix via broadcasting, and a whale list (top 1% of users by usage).

Predict before you run

  1. Will mean feature-total usage sit above or below the median? (Whales pull the mean.)
  2. What shape must the row-mean have so (users, features) / (users, 1) broadcasts?
  3. Will the top 1% of users account for closer to 1% of usage, or a lot more?

Task

Work in starter.py. Run from the repo root:

python exercises/ml/week-01/starter.py

Use the real files — the lesson's DAU picture is 7 days × 4 regions.

1. Feature ranking. Load feature_usage.csv. For each feature_name, compute total usage_count with a group-by, then convert the totals to a NumPy array and print mean / median / p90 of those feature totals.

2. Broadcasting on a real pivot. Pivot a sample of users × features into a 2-D usage matrix (fillna(0)). Divide each row by that row's mean (user-normalized usage). Shapes: (users, features) / (users, 1).

3. Whale hunt. Per user_id, sum usage. List user ids in the top 1%. How many are they? What share of all usage do they account for?

💡 Hint — row-normalize with broadcasting 
mat = pivot.to_numpy()
row_means = mat.mean(axis=1, keepdims=True)  # shape (n_users, 1)
normalized = mat / np.where(row_means == 0, 1, row_means)

Success criteria

  • Stats are over feature totals, not raw rows.
  • Normalized matrix shape matches the pivot; zero-mean rows did not explode.
  • Whale count and usage share are printed.

Debugging clues

  • Pivoting every user can be huge — sample.
  • Row mean 0 → divide by 1.
  • p90 of row usage is a different question than p90 of feature totals.

After you run

Pandas groups the ragged keys. NumPy does the slab math. Broadcasting is a shape contract, not a trick.

Week 1 — NumPy: Fast Math on Whole Columns