In this webinar, we will loads data from the UCI Online Retail data (http://archive.ics.uci.edu/ml/datasets/Online+Retail) and predicts which customers are likely to churn given their purchase activity.
Churn can be defined in many ways. We define churn to be no activity within a period of time (called the churn_period). Using this definition, a user/customer is said to have churned any form of activity is followed by no activity for an entire duration of time known as the churn_period (by default, we assume 30 days). The following figure better illustrates this concept.
(from our user guide: https://turi.com/learn/userguide/churn_prediction/churn-prediction.html)
We will dig deaper into the different parameters of the Churn Prediction toolkit, but let's start by loading some data!
In [1]:
# Let's import Graphlab Create and a few other libraries
import graphlab as gl
import graphlab.aggregate
import datetime
import time
In [2]:
#Data can come directly from a SQL database, for this webinar, we will load from a local copy
data = gl.SFrame("https://static.turi.com/datasets/churn-prediction/online_retail.csv")
data
Out[2]:
We need to do some cleanup first. The Invoice ID and Description columns are not going to help the model, and should be removed.
In [3]:
data = data.remove_columns(['InvoiceNo', 'Description'])
data
Out[3]:
Now we need to convert the InvoiceDate (which is a string) into a Python DateTime object
In [4]:
import dateutil
from dateutil import parser
def string_time_to_datetime(x):
import datetime
import pytz
return dateutil.parser.parse(x)
data['InvoiceDate'] = data['InvoiceDate'].apply(string_time_to_datetime)
Finally, we want to separate some users into a train/validation set, making sure the validation users are not in the training set, and creating TimeSeries objects out of them.
In [5]:
(train, valid) = gl.churn_predictor.random_split(data, user_id = 'CustomerID', fraction = 0.9, seed = 12)
train_trial = gl.TimeSeries(train, index = 'InvoiceDate')
valid_trial = gl.TimeSeries(valid, index = 'InvoiceDate')
Now we can load user information, which can be used to augment the churn prediction model.
In [6]:
userdata = gl.SFrame("https://static.turi.com/datasets/churn-prediction/online_retail_side_data_extended.csv")
userdata
Out[6]:
Let's now train the model.
First, let's observe the data, and see what the time range looks like
In [7]:
print "Start date : %s" % train_trial.min_time
print "End date : %s" % train_trial.max_time
In [8]:
# Period of inactivity that defines churn -- meaning that if a user stops purchasing
# items for 7 days, we'll consider them as having churned.
churn_period_trial = datetime.timedelta(days = 30)
# Different beginning of months
churn_boundary_aug = datetime.datetime(year = 2011, month = 8, day = 1)
churn_boundary_sep = datetime.datetime(year = 2011, month = 9, day = 1)
churn_boundary_oct = datetime.datetime(year = 2011, month = 10, day = 1)
In [9]:
model = gl.churn_predictor.create(train_trial,
user_data = userdata,
user_id='CustomerID',
churn_period = churn_period_trial,
time_boundaries = [churn_boundary_aug, churn_boundary_sep, churn_boundary_oct])
In [10]:
# Evaluate this model in October
evaluation_time = churn_boundary_oct
In [11]:
metrics = model.evaluate(valid_trial, evaluation_time, user_data = userdata)
In [12]:
print(metrics)
In [13]:
# metrics['precision_recall_curve'].show()
Here the question to ask is will they churn after a certain period of time. To validate we can see if they user has used us after that evaluation period. Voila! I was confusing it with expiration time (customer churn not usage churn)
In [14]:
# Make predictions in the future.
predictions_trial = model.predict(valid_trial, user_data = userdata)
predictions_trial.print_rows()
In [15]:
predictions_trial.sort('probability', ascending=False).print_rows(20,max_column_width=20)
In [16]:
predictions_trial.sort('probability', ascending=False)[200:300] .print_rows(20,max_column_width=20)
In [17]:
model.trained_model
Out[17]:
In [18]:
model.trained_model.get_feature_importance()
Out[18]: