ecommerce-using-bigquery-ML

Predict wich new visitors will come back and purchase using the Google Merchandise Store dataset

This repository has the purpose to apply BigQuery ML in the Google Merchandise Store dataset. The tasks doing here was:

---

Experiments

First approach

Training using a logitisc regression model using the following structure:

• Features:
  • Bounces;
  • time_on_site.
• Labels: will_buy_on_return_vistit

The file to train this models are available here: train_model1.sql. The file to run the evaluation it located here: eval_model1.sql

Second approach

Training using a logitisc regression model using the following structure:

• Features:

  • Bounces;
  • time_on_site;
  • page_views;
  • source,
  • medium,
  • channels_grouping;
  • device_category;
  • country.

• Labels: will_buy_on_return_vistit

• The file to train this models are available here: train_model2.sql.

• The file to run the evaluation it located here: eval_model2.sql

Third approach

Training using a XGBOOST model using the following structure:

• Features:

  • Bounces;
  • time_on_site;
  • page_views;
  • source,
  • medium,
  • channels_grouping;
  • device_category;
  • country.

• Labels: will_buy_on_return_vistit

• The file to train this models are available here: train_xgboost.sql.

• The file to run the evaluation it located here: eval_xgboost.sql

---

Results

For each experiments we use the ROC (Receiver Operating Charateristics) curve where we try to maximize the area under the curve or AUC.

Example of ROC Curve definition.

Approachs	ML Model	Numbers of Independent Features	ROC AUC
First Approach	Linear Regression	2	0.78
Second Approach	Linear Regression	8	0.92
Third Approach	XGBoost	8	0.94