Predicting User Churn for a Streaming Service using Spark.

The value in column “length” represents a song’s duration and therefore is null for all page events other than “NextSong”. Using time passed between a “NextSong” event and the following event, it would be possible to calculate the time a user spent listening to a song.

In the dataset each user accesses the streaming service always from the same location. This could mean that users access the service only from home. This hypothesis is supported by the fact that in the column “user agent” there are no entries for mobile devices.

It seems as the user log of this dataset contains a period of two months as the column “ts” containing timestamps has the minimum value October 1st and maximum value December 3rd.

After getting an understanding of the data, the next step of the CRISP-DM is to prepare the dataset for the model to train.

The first step of data preparation is cleaning the data from invalid or missing data. In this project, this step includes records without user ids or session ids. After that, an exploratory data analysis will be conducted to find possible features for the customer churn prediction.

Data Cleaning

There are user Id values with empty strings. These empty user Ids appear for instance when a user enters the streaming service without logging in. All records (8346) containing an empty user Id will be dropped, resulting into 278,154 rows in the cleaned dataset. This were the users with authentification status “Logged Out” (8249) and “Guest” (97).

Goal of the following exploratory data analysis is to observe differences in the behaviour of customers who stayed versus customers who churned. One way is to explore aggregates on these two groups of users, observing how much of a specific action they experienced per a certain time unit or number of songs played.

Labeling data

First a column “churned” will be created to use as a label for differentiation between customers who churned and those who stayed with the service. This column will later be used as label for training the supervised machine learning model. The “Cancellation Confirmation” events serve to define the exact moment of churn, which appear for both paid and free users.

The table below shows an example of the last six user’s interactions of the user called Adriel before churning. After listening to five songs, Adriel downgrades a song, then enters the cancel page and deletes his account.

+---------+-------------------------+--------------------+
|firstName|    page                 |              artist|
+---------+-------------------------+--------------------+
|   Adriel|NextSong                 |               Tonic|
|   Adriel|NextSong                 |          Arch Enemy|
|   Adriel|NextSong                 |Les Ogres De Barback|
|   Adriel|NextSong                 |The Notorious B.I.G.|
|   Adriel|NextSong                 |          Nickelback|
|   Adriel|Downgrade                |                None|
|   Adriel|Cancel                   |                None|
|   Adriel|Cancellation Confirmation|                None|
+---------+-------------------------+--------------------+

Check for imbalance in Data

A new column “cancellation_event” is created to mark the exact event of cancellation confirmation. With the new column it is possible to check, if the dataset is balanced regarding the number of users who eventually churn and those who stay. This is interesting, because of the users amount who churn is substantially lower than those who will not churn. Then, there will be less examples for the model to train how these users behave.

There is an imbalance in the dataset regarding the amount of users who churned versus those who stayed. Of a total 225 users, only 52 users eventually churned, which equals a churn rate of 23.11% .

How does this imbalance scale in amount of users on the amount of interactions? Only 16.13% of user interactions are by churned customers. The amount of data available regarding interactions to analyse the difference in behaviour for users who stayed versus users who churned is clearly imbalanced.

Imbalance in the training data can lead to naive behaviour in the prediction of the supervised machine learning model. With 76.89% of users not churning a prediction accuracy of 76.89% can be achieved by simply always predicting “not churned” [3].

There are different ways to handle imbalanced data before feeding it into machine learning algorithms. One way would be to manipulate the input data by either undersampling data of loyal users, oversampling the data of churned users or generating synthetic data. In this project the way to handle the imbalance in data will be by creating additional features.

Feature Creation

The next step is Feature creation for the machine learning model. Features are created from the available data with the goal to allow the model to distinguish between users who churn and those who not. An useful feature exposes differences in the behaviour of loyal users and those who probably churn.

One example for a possible feature is the time since registration. The mean duration from registration to last interaction with the streaming service is 57.8 days for users who eventually churn and 87.1 days for users who stay with the service. It seems intuitive that loyal users in average stay longer with the service.

              +-------+----------------------------+
|churned|avg(days_since_registration)|
+-------+----------------------------+
|   true|           57.80769230769231|
|  false|           87.14450867052022|
+-------+----------------------------+

The violin plot below shows that this difference in mean is also visible in the distribution of the values for users who churned versus users who did not churn.