A pipeline that turns ad logs into features for training a CTR model. This is the problem closest to an ads ML engineer's actual work, and it confronts data leakage head-on.

Input

rows = [
    {"user": "u1", "ad": "a1", "device": "ios",     "click": 1},
    {"user": "u1", "ad": "a2", "device": "ios",     "click": 0},
    {"user": "u2", "ad": "a1", "device": "android", "click": 1},
]

Each row is one impression. click is 0 or 1.

Features to build

Three features for each of the three axes (user, ad, device) — nine in total, appended to the original row.

{axis}_impressions   impressions for that axis value
{axis}_clicks        clicks for that axis value
{axis}_ctr           clicks / impressions   (0.0 when the denominator is 0)

Shared rules

• Keep the original keys (user, ad, device, click) and add the nine.
• Preserve the input row order exactly.
• Round CTR to 6 decimal places (round(x, 6)).
• Do not mutate the input dictionaries — build and return new ones.
• It must work on values never seen before (there is no predefined vocabulary).

Level structure

Level 2 is the heart of this problem.

Implement build_ctr_features(rows).

Each row's statistics count every impression where that axis value appears anywhere in the dataset.

rows = [
    {"user": "u1", "ad": "a1", "device": "ios",     "click": 1},
    {"user": "u1", "ad": "a2", "device": "ios",     "click": 0},
    {"user": "u2", "ad": "a1", "device": "android", "click": 1},
]
build_ctr_features(rows)[0]
# {"user": "u1", "ad": "a1", "device": "ios", "click": 1,
#  "user_impressions": 2, "user_clicks": 1, "user_ctr": 0.5,
#  "ad_impressions": 2,   "ad_clicks": 2,   "ad_ctr": 1.0,
#  "device_impressions": 2, "device_clicks": 1, "device_ctr": 0.5}

Two passes are enough — one to aggregate, one to attach.