Logistic regression codelab

1. Intro

This codelab will teach you how to use logistic regression to understand the degree to which features such as gender, age group, impression time, and browser type correlate to a user's likelihood to click an ad.

Prerequisites

To complete this codelab, you'll need enough high quality campaign data to create a model.

2. Pick a campaign

Begin by selecting an old campaign containing a large quantity of high quality data. If you don't know which campaign is likely to have the best data, run the following query on the oldest full month of data that you have access to:

SELECT
  campaign_id,
  COUNT(DISTINCT user_id) AS user_count,
  COUNT(*) AS impression_count
FROM adh.google_ads_impressions

ORDER BY user_count DESC;

Selecting data that is 12-13 months old allows you to train and test your model on data that will soon be removed from Ads Data Hub. If you encounter model-training limits on this data, those limits will end when the data is deleted.

If your campaign is particularly active, a week of data may be enough. Lastly, the number of distinct users should be 100,000 or more, especially if you're training using many features.

3. Create a temporary table

Once you've identified the campaign you'll use to train your model, run the query below.

CREATE TABLE
 binary_logistic_regression_example_data
AS(
 WITH all_data AS (
   SELECT
     imp.user_id as user_id,
     ROW_NUMBER() OVER(PARTITION BY imp.user_id) AS rowIdx,
     imp.browser as browser_name,
     gender_name as gender_name,
     age_group_name as age_group_name,
     DATETIME(TIMESTAMP_MICROS(
       imp.query_id.time_usec), "America/Los_Angeles") as impression_time,
     CASE # Binary classification of clicks simplifies model weight interpretation
        WHEN clk.click_id.time_usec IS NULL THEN 0
        ELSE 1
     END AS label
   FROM adh.google_ads_impressions imp
     LEFT JOIN adh.google_ads_clicks clk USING (impression_id)
     LEFT JOIN adh.gender ON demographics.gender = gender_id
     LEFT JOIN adh.age_group ON demographics.age_group = age_group_id
   WHERE
     campaign_id IN (YOUR_CID_HERE)
 )
 SELECT
   label,
   browser_name,
   gender_name,
   age_group_name,
   # Although BQML could divide impression_time into several useful variables on
   # its own, it may attempt to divide it into too many features. As a best
   # practice extract the variables that you think will be most helpful.
   # The output of impression_time is a number, but we care about it as a
   # category, so we cast it to a string.
   CAST(EXTRACT(DAYOFWEEK FROM impression_time) AS STRING) AS day_of_week,
   # Comment out the previous line if training on a single week of data
   CAST(EXTRACT(HOUR FROM impression_time) AS STRING) AS hour,
 FROM
   all_data
 WHERE
   rowIdx = 1 # This ensures that there's only 1 row per user.
   AND
   gender_name IS NOT NULL
   AND
   age_group_name IS NOT NULL
);

4. Create and train a model

It's a best practice to separate your table creation steps from your model creation steps.

Run the following query on the temporary table you created in the previous step. Don't worry about providing start and end dates, as these will be inferred based on data in the temporary table.

CREATE OR REPLACE
MODEL `binary_logistic_example`
OPTIONS(
   model_type = 'adh_logistic_regression'
)
AS (
   SELECT *
   FROM
       tmp.binary_logistic_regression_example_data
);

SELECT * FROM ML.EVALUATE(MODEL `binary_logistic_example`)

5. Interpret results

When the query finishes running, you'll get a table that resembling the one below. Results from your campaign will differ.

Row

precision

recall

accuracy

f1_score

log_loss

roc_auc

1

0.53083894341399718

0.28427804486705865

0.54530547622568992

0.370267971696336

0.68728232223722974

0.55236263736263735

Examine weights

Run the following query to look at the weights to see what features contribute to your model's likelihood to predict a click:

SELECT * FROM ML.WEIGHTS(MODEL `binary_logistic_example`)

The query will produce results similar to those below. Note that BigQuery will sort the given labels and choose the "smallest" to be 0 and the largest to be 1. In this example, clicked is 0 and not_clicked is 1. Thus, interpret larger weights as an indication that the feature makes clicks less likely. Additionally, day 1 corresponds to Sunday.

processed_input

weight

category_weights.category

category_weights.weight

1

INTERCEPT

-0.0067900886484743364

2

browser_name

null

unknown 0.78205563068099249

Opera 0.097073700069504443

Dalvik -0.75233190448454246

Edge 0.026672464688442348

Silk -0.72539916969348706

Other -0.10317444840919325

Samsung Browser 0.49861066525009368

Yandex 1.3322608977581121

IE -0.44170947381475295

Firefox -0.10372609461557714

Chrome 0.069115931084794066

Safari 0.10931362123676475

3

day_of_week

null

7 0.051780350639992277

6 -0.098905011477176716

4 -0.092395178188358462

5 -0.010693625983554155

3 -0.047629987110766638

1 -0.0067030673140933122

2 0.061739400111810727

4

hour

null

15 -0.12081420778273

16 -0.14670467657779182

1 0.036118460001355934

10 -0.022111985303061014

3 0.10146297241339688

8 0.00032334907570882464

12 -0.092819888101463813

19 -0.12158349523248162

2 0.27252001951689164

4 0.1389215333278028

18 -0.13202189122418825

5 0.030387010564142392

22 0.0085803647602565782

13 -0.070696534712732753

14 -0.0912853928925844

9 -0.017888651719350213

23 0.10216569641652029

11 -0.053494611827240059

20 -0.10800180853273429

21 -0.070702105471528345

0 0.011735200996326559

6 0.016581239381563598

17 -0.15602138949559918

7 0.024077394387953525

5

age_group_name

null

45-54 -0.013192901125032637

65+ 0.035681341407469279

25-34 -0.044038102549733116

18-24 -0.041488170110836373

unknown 0.025466344709472313

35-44 0.01582412778809188

55-64 -0.004832373590628946

6

gender_name

null

male 0.061475274448403977

unknown 0.46660611583398443

female -0.13635601771194916