Classification of temporal instances

Question

Problem description

I am working for a telecom domain project where I am tasked to predict whether a customer would dispute his/her monthly bill or not.

I have following historical data elements available at customer+bill month level:

1. Customer id : Unique for each customer
2. Bill month : Feb-2015, Mar-2015 etc.
3. Bill amount : Amount in $
4. New order : Flag indicating if there were new orders during the month
5. Disputed flag : Flag indicating
6. Demographics data

Objective - during every billing cycle, upon bill generation, predict if the customer is going to dispute it or not.

Modeling considerations:

1. Customers who have raised disputes in the past are more likely to raise disputes in the future
2. Historical bill amount patterns could be good predictors of disputes
3. New order and respective bill amount variation interactions could be good predictors of disputes

Design approaches :

Standard classification algorithms (GLM/RF etc) keeping data at customer+month level itself and create appropriate features as per my hypothesis.

But I have a feeling that this is not the best approach as the account level info will be highly duplicated. The classifier would treat multiple records for a particular account across months to be independent, which is not the case.

I have been looking at papers on temporal classifications and multi-instance learning. But none of the designs I have read so far would cater to the pertinent requirements.

Any solution suggestions/ research recommendations would be of huge help!

Answer 1

I would suggest a logistic algorithm. I´ve applied once and it work very well. However, the success depends a lot on the dataset you have.

I would also apply a feature selection first in order remove redundant features

All the best

Richard Sutherland