Predict Visitor Purchases with a Classification Model in BQML

• Overview
• Explore ecommerce data
• Identify an objective
• Select features and create your training dataset
• Select a BQML model type and specify options
• Evaluate classification model performance
  • Select your performance criteria
• Improve model performance with Feature Engineering
• Results

Overview

BigQuery Machine Learning (BQML, product in beta) is a new feature in BigQuery where data analysts can create, train, evaluate, and predict with machine learning models with minimal coding.

There is a newly available ecommerce dataset that has millions of Google Analytics records for the Google Merchandise Store loaded into BigQuery. In this lab you will use this data to run some typical queries that businesses would want to know about their customers’ purchasing habits.

Explore ecommerce data

Scenario: Your data analyst team exported the Google Analytics logs for an ecommerce website into BigQuery and created a new table of all the raw ecommerce visitor session data for you to explore. Using this data, you’ll try to answer a few questions.

Question: Out of the total visitors who visited our website, what % made a purchase?

#standardSQL
WITH visitors AS(
SELECT
COUNT(DISTINCT fullVisitorId) AS total_visitors
FROM `data-to-insights.ecommerce.web_analytics`
),
purchasers AS(
SELECT
COUNT(DISTINCT fullVisitorId) AS total_purchasers
FROM `data-to-insights.ecommerce.web_analytics`
WHERE totals.transactions IS NOT NULL
)
SELECT
  total_visitors,
  total_purchasers,
  total_purchasers / total_visitors AS conversion_rate
FROM visitors, purchasers

The result: 2.69%

Question: What are the top 5 selling products?

SELECT
  p.v2ProductName,
  p.v2ProductCategory,
  SUM(p.productQuantity) AS units_sold,
  ROUND(SUM(p.localProductRevenue/1000000),2) AS revenue
FROM `data-to-insights.ecommerce.web_analytics`,
UNNEST(hits) AS h,
UNNEST(h.product) AS p
GROUP BY 1, 2
ORDER BY revenue DESC
LIMIT 5;

The result:

Row	v2ProductName	v2ProductCategory	units_sold	revenue
1	Nest® Learning Thermostat 3rd Gen-USA - Stainless Steel	Nest-USA	17651	870976.95
2	Nest® Cam Outdoor Security Camera - USA	Nest-USA	16930	684034.55
3	Nest® Cam Indoor Security Camera - USA	Nest-USA	14155	548104.47
4	Nest® Protect Smoke + CO White Wired Alarm-USA	Nest-USA	6394	178937.6
5	Nest® Protect Smoke + CO White Battery Alarm-USA	Nest-USA	6340	178572.4

Question: How many visitors bought on subsequent visits to the website?

# visitors who bought on a return visit (could have bought on first as well
WITH all_visitor_stats AS (
SELECT
  fullvisitorid, # 741,721 unique visitors
  IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid
)
SELECT
  COUNT(DISTINCT fullvisitorid) AS total_visitors,
  will_buy_on_return_visit
FROM all_visitor_stats
GROUP BY will_buy_on_return_visit

The results:

Row	total_visitors	will_buy_on_return_visit
1	729848	0
2	11873	1

Analyzing the results, you can see that (11873 / 741721) = 1.6% of total visitors will return and purchase from the website. This includes the subset of visitors who bought on their very first session and then came back and bought again.

Question: What are some of the reasons a typical ecommerce customer will browse but not buy until a later visit?

Answer: Although there is no one right answer, one popular reason is comparison shopping between different ecommerce sites before ultimately making a purchase decision. This is very common for luxury goods where significant up-front research and comparison is required by the customer before deciding (think car purchases) but also true to a lesser extent for the merchandise on this site (t-shirts, accessories, etc).

In the world of online marketing, identifying and marketing to these future customers based on the characteristics of their first visit will increase conversion rates and reduce the outflow to competitor sites.

Identify an objective

Now you will create a Machine Learning model in BigQuery to predict whether or not a new user is likely to purchase in the future. Identifying these high-value users can help your marketing team to target them with special promotions and ad campaigns to ensure a conversion while they comparison shop between visits to your ecommerce site.

Select features and create your training dataset

Google Analytics captures a wide variety of dimensions and measures about a user’s visit on this ecommerce website. Browse the complete list of fields here and then preview the demo dataset to find useful features that will help a machine learning model understand the relationship between data about a visitor’s first time on your website and whether they will return and make a purchase.

Your team decides to test whether these two fields are good inputs for your classification model:

• totals.bounces (whether the visitor left the website immediately)
• totals.timeOnSite (how long the visitor was on our website)

Question: What are the risks of only using the above two fields?

Answer: Machine learning is only as good as the training data that is fed into it. If there isn’t enough information for the model to determine and learn the relationship between your input features and your label (in this case, whether the visitor bought in the future) then you will not have an accurate model. While training a model on just these two fields is a start, you will see if they’re good enough to produce an accurate model.

SELECT
  * EXCEPT(fullVisitorId)
FROM
  # features
  (SELECT
    fullVisitorId,
    IFNULL(totals.bounces, 0) AS bounces,
    IFNULL(totals.timeOnSite, 0) AS time_on_site
  FROM
    `data-to-insights.ecommerce.web_analytics`
  WHERE
    totals.newVisits = 1)
  JOIN
  (SELECT
    fullvisitorid,
    IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM
      `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid)
  USING (fullVisitorId)
ORDER BY time_on_site DESC
LIMIT 10;

Results:

Row	bounces	time_on_site	will_buy_on_return_visit
1	0	15047	0
2	0	12136	0
3	0	11201	0
4	0	10046	0
5	0	9974	0
6	0	9564	0
7	0	9520	0
8	0	9275	1
9	0	9138	0
10	0	8872	0

Question: Which fields are the input features and the label?

Answer The inputs are bounces and time_on_site. The label is will_buy_on_return_visit.

Question: Which two fields are known after a visitor’s first session?

Answer: bounces and time_on_site are known after a visitor’s first session.

Question: Which field isn’t known until later in the future?

Answer: will_buy_on_return_visit is not known after the first visit. Again, you’re predicting for a subset of users who returned to your website and purchased. Since you don’t know the future at prediction time, you cannot say with certainty whether a new visitor come back and purchase. The value of building an ML model is to get the probability of future purchase based on the data gleaned about their first session.

Question: Looking at the initial data results, do you think time_on_site and bounces will be a good indicator of whether the user will return and purchase or not?

Answer: It’s often too early to tell before training and evaluating the model, but at first glance out of the top 10 time_on_site, only 1 customer returned to buy, which isn’t very promising. Let’s see how well the model does.

Select a BQML model type and specify options

Now that you have your initial features selected, you are now ready to create your first ML model in BigQuery.

There are the two model types to choose from:

Model	Model Type	Label Data type	Example
Forecasting	linear_reg	Numeric value (typically an integer or floating point)	Forecast sales figures for next year given historical sales data.
Classification	logistic_reg	0 or 1 for binary classification	Classify an email as spam or not spam given the context.

Which model type should you choose?

Since you are bucketing visitors into “will buy in future” or “won’t buy in future”, use logistic_reg in a classification model.

The following query creates a model and specifies model options. Run this query to train your model:

CREATE OR REPLACE MODEL `ecommerce.classification_model`
OPTIONS
(
model_type='logistic_reg',
labels = ['will_buy_on_return_visit']
)
AS
#standardSQL
SELECT
  * EXCEPT(fullVisitorId)
FROM
  # features
  (SELECT
    fullVisitorId,
    IFNULL(totals.bounces, 0) AS bounces,
    IFNULL(totals.timeOnSite, 0) AS time_on_site
  FROM
    `data-to-insights.ecommerce.web_analytics`
  WHERE
    totals.newVisits = 1
    AND date BETWEEN '20160801' AND '20170430') # train on first 9 months
  JOIN
  (SELECT
    fullvisitorid,
    IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM
      `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid)
  USING (fullVisitorId)
;

Next, you evaluate the performance of the model against new unseen evaluation data.

Evaluate classification model performance

Select your performance criteria

For classification problems in ML, you want to minimize the False Positive Rate (predict that the user will return and purchase and they don’t) and maximize the True Positive Rate (predict that the user will return and purchase and they do).

This relationship is visualized with a ROC (Receiver Operating Characteristic) curve like the one shown here, where you try to maximize the area under the curve or AUC:

In BQML, roc_auc is simply a queryable field when evaluating your trained ML model.

Now that training is complete, run this query to evaluate how well the model performs using ML.EVALUATE:

SELECT
  roc_auc,
  CASE
    WHEN roc_auc > .9 THEN 'good'
    WHEN roc_auc > .8 THEN 'fair'
    WHEN roc_auc > .7 THEN 'decent'
    WHEN roc_auc > .6 THEN 'not great'
  ELSE 'poor' END AS model_quality
FROM
  ML.EVALUATE(MODEL ecommerce.classification_model,  (
SELECT
  * EXCEPT(fullVisitorId)
FROM
  # features
  (SELECT
    fullVisitorId,
    IFNULL(totals.bounces, 0) AS bounces,
    IFNULL(totals.timeOnSite, 0) AS time_on_site
  FROM
    `data-to-insights.ecommerce.web_analytics`
  WHERE
    totals.newVisits = 1
    AND date BETWEEN '20170501' AND '20170630') # eval on 2 months
  JOIN
  (SELECT
    fullvisitorid,
    IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM
      `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid)
  USING (fullVisitorId)
));

You should see the following result:

Row	roc_auc	model_quality
1	0.7238561438561438	decent

After evaluating your model you get a roc_auc of 0.72, which shows the model has decent, but not great, predictive power. Since the goal is to get the area under the curve as close to 1.0 as possible, there is room for improvement.

Improve model performance with Feature Engineering

As was hinted at earlier, there are many more features in the dataset that may help the model better understand the relationship between a visitor’s first session and the likelihood that they will purchase on a subsequent visit.

Add some new features and create a second machine learning model called classification_model_2:

• How far the visitor got in the checkout process on their first visit
• Where the visitor came from (traffic source: organic search, referring site etc..)
• Device category (mobile, tablet, desktop)
• Geographic information (country)

Create this second model by clicking on COMPOSE NEW QUERY:

CREATE OR REPLACE MODEL `ecommerce.classification_model_2`
OPTIONS
  (model_type='logistic_reg', labels = ['will_buy_on_return_visit']) AS
WITH all_visitor_stats AS (
SELECT
  fullvisitorid,
  IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid
)
# add in new features
SELECT * EXCEPT(unique_session_id) FROM (
  SELECT
      CONCAT(fullvisitorid, CAST(visitId AS STRING)) AS unique_session_id,
      # labels
      will_buy_on_return_visit,
      MAX(CAST(h.eCommerceAction.action_type AS INT64)) AS latest_ecommerce_progress,
      # behavior on the site
      IFNULL(totals.bounces, 0) AS bounces,
      IFNULL(totals.timeOnSite, 0) AS time_on_site,
      IFNULL(totals.pageviews, 0) AS pageviews,
      # where the visitor came from
      trafficSource.source,
      trafficSource.medium,
      channelGrouping,
      # mobile or desktop
      device.deviceCategory,
      # geographic
      IFNULL(geoNetwork.country, "") AS country
  FROM `data-to-insights.ecommerce.web_analytics`,
     UNNEST(hits) AS h
    JOIN all_visitor_stats USING(fullvisitorid)
  WHERE 1=1
    # only predict for new visits
    AND totals.newVisits = 1
    AND date BETWEEN '20160801' AND '20170430' # train 9 months
  GROUP BY
  unique_session_id,
  will_buy_on_return_visit,
  bounces,
  time_on_site,
  totals.pageviews,
  trafficSource.source,
  trafficSource.medium,
  channelGrouping,
  device.deviceCategory,
  country
);

A new key feature that was added to the training dataset query is the maximum checkout progress each visitor reached in their session, which is recorded in the field hits.eCommerceAction.action_type. If you search for that field in the field definitions you will see the field mapping of 6 = Completed Purchase.

The web analytics dataset has nested and repeated fields like ARRAYS which need to broken apart into separate rows in your dataset. This is accomplished by using the UNNEST() function, which you can see in the above query.

Evaluate this new model to see if there is better predictive power:

#standardSQL
SELECT
  roc_auc,
  CASE
    WHEN roc_auc > .9 THEN 'good'
    WHEN roc_auc > .8 THEN 'fair'
    WHEN roc_auc > .7 THEN 'decent'
    WHEN roc_auc > .6 THEN 'not great'
  ELSE 'poor' END AS model_quality
FROM
  ML.EVALUATE(MODEL ecommerce.classification_model_2,  (
WITH all_visitor_stats AS (
SELECT
  fullvisitorid,
  IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid
)
# add in new features
SELECT * EXCEPT(unique_session_id) FROM (
  SELECT
      CONCAT(fullvisitorid, CAST(visitId AS STRING)) AS unique_session_id,
      # labels
      will_buy_on_return_visit,
      MAX(CAST(h.eCommerceAction.action_type AS INT64)) AS latest_ecommerce_progress,
      # behavior on the site
      IFNULL(totals.bounces, 0) AS bounces,
      IFNULL(totals.timeOnSite, 0) AS time_on_site,
      totals.pageviews,
      # where the visitor came from
      trafficSource.source,
      trafficSource.medium,
      channelGrouping,
      # mobile or desktop
      device.deviceCategory,
      # geographic
      IFNULL(geoNetwork.country, "") AS country
  FROM `data-to-insights.ecommerce.web_analytics`,
     UNNEST(hits) AS h
    JOIN all_visitor_stats USING(fullvisitorid)
  WHERE 1=1
    # only predict for new visits
    AND totals.newVisits = 1
    AND date BETWEEN '20170501' AND '20170630' # eval 2 months
  GROUP BY
  unique_session_id,
  will_buy_on_return_visit,
  bounces,
  time_on_site,
  totals.pageviews,
  trafficSource.source,
  trafficSource.medium,
  channelGrouping,
  device.deviceCategory,
  country
)
));

(Output)

Row	roc_auc	model_quality
1	0.9094875124875125	good

With this new model you now get a roc_auc of 0.91 which is significantly better than the first model.

Now that you have a trained model, time to make some predictions.Predict which new visitors will come back and purchase

Next you will write a query to predict which new visitors will come back and make a purchase.

The prediction query below uses the improved classification model to predict the probability that a first-time visitor to the Google Merchandise Store will make a purchase in a later visit:

SELECT
*
FROM
  ml.PREDICT(MODEL `ecommerce.classification_model_2`,
   (
WITH all_visitor_stats AS (
SELECT
  fullvisitorid,
  IF(COUNTIF(totals.transactions > 0 AND totals.newVisits IS NULL) > 0, 1, 0) AS will_buy_on_return_visit
  FROM `data-to-insights.ecommerce.web_analytics`
  GROUP BY fullvisitorid
)
  SELECT
      CONCAT(fullvisitorid, '-',CAST(visitId AS STRING)) AS unique_session_id,
      # labels
      will_buy_on_return_visit,
      MAX(CAST(h.eCommerceAction.action_type AS INT64)) AS latest_ecommerce_progress,
      # behavior on the site
      IFNULL(totals.bounces, 0) AS bounces,
      IFNULL(totals.timeOnSite, 0) AS time_on_site,
      totals.pageviews,
      # where the visitor came from
      trafficSource.source,
      trafficSource.medium,
      channelGrouping,
      # mobile or desktop
      device.deviceCategory,
      # geographic
      IFNULL(geoNetwork.country, "") AS country
  FROM `data-to-insights.ecommerce.web_analytics`,
     UNNEST(hits) AS h
    JOIN all_visitor_stats USING(fullvisitorid)
  WHERE
    # only predict for new visits
    totals.newVisits = 1
    AND date BETWEEN '20170701' AND '20170801' # test 1 month
  GROUP BY
  unique_session_id,
  will_buy_on_return_visit,
  bounces,
  time_on_site,
  totals.pageviews,
  trafficSource.source,
  trafficSource.medium,
  channelGrouping,
  device.deviceCategory,
  country
)
)
ORDER BY
  predicted_will_buy_on_return_visit DESC;

The predictions are made on the last 1 month (out of 12 months) of the dataset.Your model now outputs its predictions for those July 2017 ecommerce sessions. You can see three newly added fields:

• predicted_will_buy_on_return_visit: whether the model thinks the visitor will buy later (1 = yes)
• predicted_will_buy_on_return_visit_probs.label: the binary classifier for yes / no
• predicted_will_buy_on_return_visit.probs.prob: the confidence the model has in it’s prediction (1 = 100%)

Results

• Of the top 6% of first-time visitors (sorted in decreasing order of predicted probability), more than 6% make a purchase in a later visit.
• These users represent nearly 50% of all first-time visitors who make a purchase in a later visit.
• Overall, only 0.7% of first-time visitors make a purchase in a later visit.
• Targeting the top 6% of first-time increases marketing ROI by 9x vs targeting them all!