You might have heard of the two music streaming giants: Apple Music and Spotify. Which is better than the other? Well, that depends on multiple factors, like the UI/UX of their app, the frequency of new content, user-curated playlists and subscribers count. The factor which we are studying is called the churn rate. Churn rate has a direct impact on the subscribers' count and also the long term growth of the business.
So what is this churn rate anyway?
For a business, the churn rate is a measure of the number of customers leaving the service or downgrading their subscription plan within a given period of time.
Imagine you are working on the data team for a popular digital music service similar to Spotify or Pandora. Millions of users stream their favourite songs to your service every day either using the free tier that plays advertisements between the songs or using the premium subscription model where they stream music at free but pay a monthly flat rate. Users can upgrade, downgrade or cancel their service at any time so it is crucial to make sure your users love the service.
In this project, our aim is to identify the customer churn for Sparkify (a Spotify-like fictional music streaming service). This does not include the variety (like the genre, curated playlists, top regional charts) of music that the service provides. It mainly explores user behaviour and how we can identify the possibility that a user will churn. Such customers are those who decide to downgrade their service, i.e. going from paid subscription to free, or entirely leaving the service.
Our target variable is isChurn. It cannot be interpreted directly from the JSON
file, but we will use feature
engineering to create it. isChurn column is 1 for users who visited the
Cancellation Confirmation page and 0 otherwise.
Out of 225 unique users, only 52 churned (23.5%). So, accuracy will not be a good metric to handle this imbalance. We will instead use F1-Score to evaluate our model.
F1-Score is the harmonic mean of precision and recall.
Precision = True Positive / (True Positive + False Positive)
Recall = True Positive / (True Positive + False Negative)
Name of the input data file is mini_sparkify_event_data.json in data directory. The shape of our feature space is 286500 rows and 18 columns. data/metadata.xlsx contains information about the features.
A preview of data:
I have used toPandas() method above because 18 columns cannot be displayed in
a user-friendly way by PySpark's built-in .show() method. Still, if you like
to see what it looks like with .show() method, here you go (you will not like it).
-
Distribution of pages
Cancellations are less. That is what we have to predict.
We will remove the starting classes:
CancelandCancellation Confirmationin our modelling section to avoid lookahead bias.Most commonly browsed pages include activities like the addition to a playlist, home page, and thumbs up.
I have not inclued
NextPagecategory because it is present around 80% of the time and will highly skew the distribution. -
Distribution of levels (free or paid)
70% of churned users are paying customers. Customer retention is relatively more important for paid ones because they are directly connected with the revenue of the company.
-
Song length
No additional information is available from this. It just shows that most songs are 4 minutes long.
-
What type of device user is streaming from?
This is what we have expected. Windows is the most used platform.
-
Gender distribution
Males are more in number.
-
Distribution of pages based on churn
No strong conclusion can be drawn from this graph. It also shows the same common actions as seen from the univariate plot here
-
Distribution of hour based on churn
We can see that non-churn users are more active during day time.
-
Behaviour across weekdays
Activity is more on weekdays. Especially for churned users. However, this change is not significant.
-
Behaviour at the month level
Non-churn users are generally less active at the start of the month as compared to churn users, and the opposite is the case at the EOM.
For page column, we have 22 distinct values:
- About
- Add Friend
- Add to Playlist
- Cancel
- Cancellation Confirmation
- Downgrade
- Error
- Help
- Home
- Login
- Logout
- NextSong
- Register
- Roll Advert
- Save Settings
- Settings
- Submit Downgrade
- Submit Registration
- Submit Upgrade
- Thumbs Down
- Thumbs Up
- Upgrade
We are interested in the fifth category. Our target variable is
isChurn. It is 1 if the user visited Cancellation Confirmation page and
0 otherwise.
This imbalanced data suggests that we should not use accuracy as our evaluation metric. We will use F1 score instead and use under-sampling to further optimise it.
- Handling null values
First, we will remove null values for some columns. There are two distinct number of null values observed: 8346 and 58392.
58392 is 20% of the data (286500) and 8346 is merely 2%. So we keep the columns which
have 2% nans and see for the 20% one's whether we can impute the missing
values in some way.
These are the columns with 20% missing values. Seems like it is difficult to impute them. We will drop the respective rows with null values for these columns.
We have the same training and testing features for all the models. PySpark's ML
library(pyspark.ml) has access to the most common machine learning classification
algorithms. Others are still in development, like
Tree Predictions in Gradient Boosting to Improve Prediction Accuracy.
The ones which we'll be using are:
Since the class distribution is highly imbalanced, we will perform random undersampling to optimize our F1 score.
F1 is the harmonic mean of precision and recall. Precision and recall are calculated in the following way:
This article will deepen your understanding on why to use F1 score when evaluating your model on imbalanced data set.
Comparison of average metrics before and after undersampling.
| Model | Average Metrics Before | Average Metrics After |
|---|---|---|
| Logistic Regression | 0.717, 0.684, 0.681 | 0.486, 0.344, 0.192 |
| Random Forest Classifier | 0.710, 0.699, 0.698 | 0.540, 0.537, 0.499 |
| Gradient Boosting Tree Classifier | 0.710, 0.705, 0.684 | 0.629, 0.627, 0.616 |
GBTClassifier provided a fairly good F1 score of 0.64 after undersampling.
I enjoyed the data pre-processing part of the project. For data visualisation part, instead of using “for” loops to get arrays for our bar chart, I first converted our Spark data frame to Pandas data frame using toPandas()method. Data visualisation is easier from then on.
The shape of our final data is just 225 x 32. This is too small to generalise our model. Just 225 users for a streaming company? That’s nothing. 12 GB data might provide some useful results. If you want more statistically significant results then I suggest you run this notebook on Amazon EMR for the 12 GB dataset. I have skipped that part for now because it costs $30 for one week.
Some of the challenges which I faced in this project are:
- Official documentation of PySpark as compared to Pandas
- For the sake of mastering Spark, we only used the most common machine learning classification models instead of using the advanced ones
- Highly imbalanced data led to a poor F1 score
- If you run this notebook on your local machine without any change, then it will take around an hour to run completely
-
Folders gbtModel, lrModel and rfModel
Saved models before under-sampling
-
Folders new_gbt_model, new_lr_model and new_rf_model
Saved models after under-sampling
-
helper.py
Helper functions for Plotly visualisations
This project uses Python 3.6.6 and the necessary libraries are mentioned in requirements.txt file.
Thanks to Udacity for providing such a challenging project to work on. At first, I was vexed with functional programming in Python, but the instructors were very clear in their approach. Now I am looking for more projects that are build atop Spark.