The Customer Datathon

Bitphy and Ubiqum Code Academy are jointly hosting a datathon aimed at data analysis students. Our goal is to use data analysis and machine learning to understand and predict customer behaviour at retail shops.

The datathon will take place from 9am to 7pm on March 14th, 2019 at a joint location of Ubiqum's offices in Barcelona, Berlin and Amsterdam.

We have provided a dataset containing anonymized transactions carried out by individual identified customers from 25 brick-and-mortar retail shops specialised in groceries in the province of Barcelona. Such transactions are dated between January, 2016 and November, 2018.

The problem: can we predict customer behaviour?

Goal

Predict, for each of the 1190 customers in the datasets/customers_val.txt file, the date of the first purchase after November 1st, 2018, and the amount of such purchase.

Validation metric

An individual customer submission will be considered correct if the predicted date matches the date of the real purchase and if the predicted billing's absolute error is at most 10€. Then we will look at accuracy: we will divide the number of correct predictions by 1190, thus obtaining a metric betweeen 0 and 1.

Available data

All datasets provided for this datathon are contained in the datasets/ folder. We briefly describe them below. For a deeper description of the available data , an initial exploration and a fitting of a dummy model are available here.

Besides the provided datasets, teams are free to use whatever open data sets they might see fit, provided they satisfy this definition.

Data for feature engineering and model training

• customers.txt: a list of the 8723 customer ids appearing in the dataset.

• X.csv: the basket-detail shopping history of customers appearing in customers.txt between the dates of January 1st, 2016 and October 31st, 2018. Shape: (2281130, 9).

• y.csv: the date and billing of the first purchase effected by all customers appearing in customers.txt after November 1st, 2018. Shape: (8723, 3).

Validation data

• customers_val.txt: a list of the 1190 customer ids which have been selected to assess model accuracy in the competition.

• X_val.csv: the basket-detail shopping history of customers between the dates of January 1st, 2016 and October 31st, 2018. Shape: (323227, 9).

How to prepare a submission

Submissions are expected to be in the form of a dataframe with three columns labelled customer, date and billing and a total of 1190 rows.

An example of a valid submission may be found in the datasets/sample_submission.csv file. Its first 5 rows look as follows:

customer	date	billing
sW2Tg	2018-11-04	20.07
jL2H_	2018-11-02	24.82
ows9Y	2018-12-20	12.41
bNtuI	2018-11-02	21.42
Bg9db	2018-11-02	14.11

Accuracy metric

A score method has been supplied in order to validate the outcome of a predictive model, and it is available as an importable module at src/validations.py. This is its signature:

def score(y, yhat):
    """Compute the score of predictions yhat by comparing to real data y.
    Both yhat and y have to be dataframes containing columns
        ['customer',
         'date',
         'billing']
    and must have exactly the same number of rows.

    A prediction for a given customer is computed as correct whenever the predicted date
    matches the real date and the predicted billing absolute error is lower or equal than
    10 €.

    :param y: (pd.DataFrame) real data.
    :param yhat: (pd.DataFrame) predicted data.
    :returns: score of the prediction.
    """

The example submission we just discussed comes from training a dummy predictor. In order to get its score, we would run the following code:

from src.validations import score

# Assuming y_real and y_pred contain the target
# values and the predictions, respectively.
accu = score(y_real, y_pred)

As a baseline for comparison, this dummy predictor attains a score of accu = 0.046218.

How to submit a prediction

A live submission tracking system has been deployed by @gparrello. This deployment uses a REST API. Gerardo has kindly provided code for quick result submission to this tracking system.

Each team will be provided with a configuration file containing a JWT (JSON Web Token) token, which will have to be passed to the submission function. An example of such a config file is available at src/config.ini. Download this file and save it in your computer.

• R users: use the submit_predictions function in the src/client.R file. An example of a code that uses submit_predictions to submit our sample submission may be found in src/example.R.

• python users: use the submit_predictions method in the module src/client.py. An example of a code that uses submit_predictions to submit our sample submission may be found in src/example.py.

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.