Updating project to work with TensorFlow 2.18

This repo is created as a fork of the Customer Lifetime Value project from google. A pull request for these changes has been created. The changes in this repo enable the use of TensorFlow 2.18.

The code has has been tested using the Ubuntu 24 operating system running Python 3.12 and the NVIDIA RTX A5000 graphics card.

List of changes

• updated package name from sklearn to scikit-learn in Setup.py

• updated notebooks from the notebook folder to:

  • use a DATA_FOLDER variable for location of input and output files.

  • replaced %%script blocks with %%bash since %%script is no longer supported.

  • added an extra dimension for y_train and y_eval during the fit call to make them 2-dimensional arrays. Without this, the zlin loss (ltv.zero_inflated_lognormal_loss) will fail as it has a checks for target variable being two dimensional.

  • removed quote characters around the company variable in calls to the pandas query function in notebooks of the kaggle_acquire_valued_shoppers_challenge folder. This may have worked in previous versions of pandas, but it silently returns an empty dataframe if a string is used as a query value against a numeric column.

  • replaced referenced to LinearModel with a Sequential linear model as this class is longer supported.

  • moved the numeric input field in kdd_cup_98/regression.ipynb to the last parameter. Due to the shape of this parameter (21,) and the presence of other features, if this paramter is not the last, TensorFlow throws an error during the call to the fit method.

• added environment.yml to save packages used in the conda environment used to build this project. This includes NVIDIA libraries.

• added requirements.txt file.

TLDR

kaggle acquire valued shoppers challenge

There are three notebooks in the folder notebooks/kaggle_acquire_valued_shoppers_challenge

• preprocess_data.ipynb has the code for processing the raw transaction file to build company specific feature files. This code is repeated in the other two notebooks. The preprocessing involves:

  • Filtering of all records. Only select transactions with positive values; this excludes all returns which have a negative value. So the label won't reflect the returns nor will the calibration value

  • Generating the calibration value. Sum up the total purchase amount for the first day of shopping.

  • Generating calibration attributes. Take the most expensive transaction and select its 'chain', 'dept', 'category', 'brand', 'productmeasure' values for the first day of shopping. Note, all other transactions are ignored. Any null values for these attributes are replaced by UNKNOWN.

  • Generating the label/holdout value. This is the total amount purchased by a customer in one year

The zero_inflated_lognormal_loss function used by both regression and classification notebooks, requires three inputs which are generated as the three output nodes of these models.

• regression.ipynb

To predict using a regression model, call the intial predict function followed by a call to the zero_inflated_lognormal_pred function, passing it all three output node values e.g.

logits = model.predict(x=x_eval, batch_size=1024)
y_pred = ltv.zero_inflated_lognormal_pred(logits).numpy().flatten()

• classification.ipynb

To predict using a classification model, call the intial predict function followed by a call to the sigmoid function passing it the values of the first of the three output node values e.g.

logits = model.predict(x=x_eval, batch_size=1024)
y_pred = K.sigmoid(logits[..., :1]).numpy().flatten()

Lifetime Value

Accurate predictions of customers’ lifetime value (LTV) given their attributes and past purchase behavior enables a more customer-centric marketing strategy. One challenge of LTV modeling is that some customers never come back, and the distribution of LTV can be heavy-tailed. The commonly used mean squared error (MSE) loss does not accommodate the significant fraction of zero value LTV from one-time purchasers and can be sensitive to extreme large LTV from top spenders.

We model the distribution of LTV given associated features as a mixture of zero point mass and lognormal distribution, which we refer to as zero-inflated lognormal (ZILN) distribution. This modeling approach enables us to capture the churn probability and account for heavy-tailedness nature of LTV at the same time, and also allows for easy uncertainty quantification of the point prediction. The proposed loss function can be used in both linear models and deep neural networks (DNN). We also advocate normalized Gini coefficients to quantify model discrimination and promote decile charts to assess model calibration.

The proposed loss function (implemented in Keras) and evaluation metrics are integrated into a python package. And we demonstrate the predictive performance of our proposed model in notebooks on two real-world public datasets.

Paper

Wang, Xiaojing, Liu, Tianqi, and Miao, Jingang. (2019). A Deep Probabilistic Model for Customer Lifetime Value Prediction. arXiv:1912.07753.

Installation

The easiest way is propably using pip:

pip install -q git+https://github.com/seyedrezamirkhani/lifetime_value

If you are using a machine without admin rights, you can do:

pip install -q git+https://github.com/seyedrezamirkhani/lifetime_value --user

If you are using Google Colab, just add "!" to the beginning:

!pip install -q git+https://github.com/seyedrezamirkhani/lifetime_value

Package works for python 3 only.

Usage

Package can be imported as

import lifetime_value as ltv

notebooks

The best way to learn how to use the package is probably by following one of the notebooks, and the recommended way of opening them is Google Colab.

Kaggle Acquire Valued Shoppers Challenge Dataset

This Kaggle challenge provides almost 350 million rows of completely anonymised transactional data from over 300,000 shoppers. We use the transactional data to demonstrate LTV modeling.

We download the transaction.csv (21GB) file from Kaggle server and prepare csv files for each of top 20 most common companies. Then we train a Keras model to predict customer's lifetime value and returning probability.

The raw data is available here.

• Data preparation downloads kaggle data transaction.csv and preprocesses the top 20 most common companies' data to customer-level one. This is optional, and running this will save time for regression and classification because the data are cached.
• Regression trains a Keras regression linear/dnn model with specified loss function and evaluates the results.
• Classification trains a Keras classification linear/dnn model with specified loss function and evaluates the results.

KDD Cup 98

The Second International Knowledge Discovery and Data Mining Tools Competition (a.k.a., the KDD Cup 1998) provides a dataset collected by Paralyzed Veterans of America (PVA), a non-profit organization that provides programs and services for US veterans with spinal cord injuries or disease. The organization raised money via direct mailing campaigns and was interested in lapsed donors: people who have stopped donating for at least 12 months. The provided dataset contains around 200K such donors who received the 1997 mailing and did not make a donation in the previous 12 months. We tackle the same task of the competition, which is to predict the donation dollar value to the 1997 mailing campaign.

The raw data is available here.

• Regression trains regression models and makes comparisons on different methods.

People

Package is created and maintained by Xiaojing Wang, Tianqi Liu, and Jingang Miao.