0

I am developing a combined-algorithm model in MS Azure Machine Learning Studio that should be able to predict whether a Telco-customer churn or not. Given that I have 19 variable features, e.g. monthly fee, usage length etc., how do I combine two useful algorithms? And how do I know these provide the highest Accuracy (Highest possible Accuracy needed), ie. which elements do I yet need to add and how should I use for predicting churn behaviour onto another dataset of a "fresh" set of customers?

I have: 1) Edit Metadata Having excluded the User_ID variable I have used the Edit Metadata element to label the Attrition variable (Attrition is whether a customer has churned, i.e. Yes or No). Simultaneously I have transformed Attrition into a categorical variable, specifying that the selected values should be treated in two categories, i.e. Yes or No.

Normalize data Since the three identified numerical variables (Usage_Length, Monthly_Fee and Total_Fee) a quite different in scale, e.g. Max(Monthly_Fee) is at 78,80 while Max(Total_Fee) is at 5.789,87, I have normalized Monthly_Fee and Total_Fee using the LogNormal Transformation method.

Edit Metadata (2nd use) Having normalized two of the numerical values, I have made all non-numerical features, e.g. User_Gender, Is_Senior etc., into categorical values to make them useful for the coming analysis.

Split data Once the above steps have been carried out, I have made a testing/training split of 0.2 and 0.8, respectively on which I run the models.

Choice of algorithm I have selected Two-class Boosed Decision Tree and Two-Class Decision Forest as they provide the highest possible individual Accuracy; 0.963 and 0.967, respectively.

No coding used - only elements added.

I expect the highest possible Accuracy, currently at 0.967 when combining the models into an Evaluation elementCurrent Model Screenshot

Christian_Peter
  • 1
  • 1
  • are you talking about Azure Machine Learning Studio? Or the Azure ML Visual Interface? Or something else entirely? – silent Jul 22 '19 at 10:31
  • Azure Machine Learning Studio – Christian_Peter Jul 22 '19 at 10:34
  • what exactly is your question, i.e. what exactly do you want to achieve? – silent Jul 22 '19 at 10:38
  • I want to develop a model based on two algorithms that is able to predict a variable (Attrition, i.e. whether a customer churn or not; Yes or No) based upon 19 other variables, e.g. for how long the customer has been with the company, how much he/she pays each month for his/her services etc. Essentially what I need is guidance on which elements to add from having Evaluated the model to be able to apply the model on a new dataset. Does it make sense? – Christian_Peter Jul 22 '19 at 10:42
  • So the main question would be as per the title: How do I combine two algorithm into one model that can predict a variable? – Christian_Peter Jul 22 '19 at 10:47
  • are the two algorithms working on different variables of your input dataset or why do you need to use both of them? Or what does combining for your look like? Average them? Take the higher value? – silent Jul 22 '19 at 11:06
  • No, they are not working on different variables. They should provide similar output in order for me to be able to compare them. I am doing an exam in which it is required to develop two models. I believe that the model I have developed (see attached picture in the question above; Current Model Screenshot) is fairly ok. Am just wondering how I make these models test another dataset? Should I do it in the same experiment or should I somehow drag it into another experiment with the new dataset? – Christian_Peter Jul 22 '19 at 12:16
  • And also wondering how do I find out which variable has the highest predictable power for whether the customer is churning or not? I cannot seem to visualize or summarize these findings in any elements? – Christian_Peter Jul 22 '19 at 12:21
  • Have you looked into Azure Automated ML? Not sure if model interpretability is possible in the same way in ML Studio. But with the Auto ML visual interface should be also fairly easy for you to build your models. https://learn.microsoft.com/en-us/azure/machine-learning/service/concept-automated-ml It brings you feature importance out of the box: https://learn.microsoft.com/en-us/azure/machine-learning/service/how-to-understand-automated-ml#explain-model – silent Jul 22 '19 at 12:41
  • Thx - much appreciated – Christian_Peter Jul 23 '19 at 06:21

0 Answers0