This project aims to build a machine learning algorithm for predicting whether customers visiting Google's Merchandise Store will make a purchase in the next 7 days.
Contributed by: Lim Si Ling Evelyn, Lee Meng Yong, Jiang Nan, Desmond Ho
Pls contact desmondhw@gmail.com for any questions.
Data:
Google Merchandise Store clickstream data from Google Bigquery from 1 November 2016 to 30 July 2017.
Dataset was extracted by incorporating SQL queries into the Python codes in Google Colab - See Data_Extraction_and_Exploratory_Data_Analysis.ipynb under codes folder.
https://bigquery.cloud.google.com/table/bigquery-public-data:google_analytics_sample.ga_sessions_20170801
Training Dataset:
Data\test_dataset.csv
Testing Dataset:
Data\train_validation_dataset.csv
Note: Target variable (transaction) is highly imbalanced with only 3.9% of the non-bounce sessions making a transaction.
Performance Metric:
F2-score. F2-score considers both precision and recall in the measurement, but implicit computation will give a higher weightage to recall measure, i.e. the ability for the model to be able to capture as many potential customers as possible.
Models:
Random Forest using data w/o any LDA feature engineering and w/o PCA obtained the highest F2-score for the test data out of all the combinations tried.
Best model
- Codes\Models without LDA _ PCA.ipynb
Other models
- Codes\Models with LDA.ipynb
- Codes\Models with PCA.ipynb
- Codes\Models without LDA _ PCA with undersampling.ipynb
Tableau:
- Codes\outliers.twbx - This visual demonstrated why Anomaly Detection models didn't work well due to the minority class(purchase) not appearing as anomalies, but clustered with the majority.
Report:
Our full report can be found in: 'Project Report - Applied Machine Learning.docx'
Requirements:
- Python 3.7
- Jupyter Notebook
- Tableau 2019.2
Python Libraries Required:
- datetime
- numpy
- pandas
- scikit-learn
- matplotlib.pyplot
- seaborn
- dmatrices
- statsmodels
- lightgbm
Python libraries above can be installed via pip. Note that we only tested with the versions above, newer versions might not work.