version: 2.0.0

Overview

In today's smartphone world, mobile games are able to attract large audience resulting large number of game studios developing mobile games.

People have large variety of games to play resulting higher churn rate and revenue loss.

Problem Statement

Predict whether a particular user will abandon the game or not.

Our Approach

To solve the problem, we use IntelliHub . Lets see how IntelliHub works:

• First user needs to uploads the train as well as test data.
• It is stored in the cloud.
• Data from cloud then goes to PhoenixML (a Machine Learning Module in IntelliHub ). In PhoenixML, you have 4 ML Libraries (Weka, H2O, TensorFlow & Scikit-learn) to choose from.
• Using PhoenixML, you train a model.
• You can evaluate using RMSE, accuracy and other different metrics.
• You can then use the above model for prediction.

Data Definition

The data set contains information about 2000 players. The variables in the data set are:

• User_id: Player unique identification number
• Achievements: Number of times a player was rewarded for successfully completing a level collecting all possible coin
• Challenges: Number of challenges completed
• No_of_sessions: Number of times a particular player opened the game
• Session_length: Total time a player played the game (in minutes)
• Coin_earned: Coins collected while playing a game
• Coin_spent: Coins spent either to buy something or get extra life
• Player_activity.Bal:Either player is churned or is active (Label)

Studio gives the flexibility of choosing features upto 20 columns.

Model Building

• We want to create a model in which it can predict either a particular player will churn in the near future or not.
• This can help to run player retention campaign for that particular player.
• We will be building classification models using two algorithms: NaiveBayesMultinomial & RandomForest on IntelliHub .

We will evaluate & compare each model and accordingly choose the best model for our test dataset.

While doing our modeling with PhoenixML we will be using some terminologies:

• Training Data : The data you use for training your model. It contains all the information you have collected about the problem statement.
• Test Data : The data you use for testing the model. You can make predictions on this data.
• Features : These are all the columns in your dataset which you use for training your model.
• Class/Label : This is the column which identifies the particular record and the one you want to predict.
• Accuracy : This is the percentage of data correctly predicted when you apply the model to your data.

First we need to create an App from console and enable API for PhoenixML.In Studio, we can build model in two ways:

• Using SDK (For Developer)
• Using Interface (For Non-Developer)

Using SDK

If you want to develop a model using SDK, just copy API key from IntelliHub Console.

Using Interface (Non-Developer)

1. Once we have created an App we need to go to the Phoenix ML.

   

2. Once we are at the Phoenix ML dashboard we can see that the App is enabled.

   

3. Before we can train our model, we need to upload the dataset in Dataset Management.

   
   Navigate to Dataset Management
   
   Uploading Train & Test dataset

4. After we have uploaded the training data, we create a new model and need to configure it.

   1. First you need to name your model.
   2. How you want to split your data like train and validation..
   3. Type of Model ie Classification or Regression. Here we are using Classification
   4. We will select Random Forest as our algorithm.
   5. Now we have to select our training dataset which we uploaded earlier.
   6. Select features from the list which you want to use for building the model. Here we are using all features except Item_Identifier and Item_Outlet_Sales.
   7. Label is our target variable that is player activity.
   8. And then click Train.
   

5. After we come back to the Phoenix ML>Explore dashboard which has a list of all the models trained.

   

6. After the model gets trained if success it will show a green TRAINED flag and if it has failed the it gets represented with a red FAILED flag. Reason for model fail can be incorrect data type in the dataset like you label (Target variable) is categorical variable while using Linear Regression.

   Same way we will create new model but this time we will use NaiveBayesMultinomial Algorithm and name the model ‘Player Churn Model - NaiveBayesMultinomial’.

   

   Once the model gets successfully completed we can see the results i.e the Pearson’s Correlation graph, Accuracy, Kappa, Error rate , Prior Entropy and ROC curve for both the models.

   
   RandomForest Model Result
   
   NaiveBayesMultinomial Model Result

7. Model Selection:

   We have built two models:

   Model	Accuracy	Kappa	Error Rate	Prior Entropy
   RandomForest	96.83%	0.911	0.032	0.806
   NaiveBayesMultinomial	56.17%%	-0.042	0.438	0.787

8. Predicting Test Data

   To predict the player activity for the test data, we need to come back to the Phoenix ML>Explore dashboard which has a list of all the models trained. We can go to the predict option. And enter the model name and the test dataset we want to use.

   We will select ‘Player Churn Model - RandomForest’.

   

9. After the successful compilation of the predictive model we will get the results. It will predict player activity as churn and active. You will also get churn probability and active probability. We can view the results in a separate window or download a csv file.

   
   Prediction Result

Summary

• This model helps us in determining set of players who will probably churn.
• We can give this players certain incentives like coins so that they can buy extra lives, unlock level,etc.
• We can also explore characteristics of this probable churns and determine the purpose of churn. Accordingly we can make changes the experience for player retention.