Optimizing Tarrif Plans for Telephone Subscribers Based on Past Usage

  • Context: Graduate 
  • Thread starter Thread starter dharavsolanki
  • Start date Start date
dharavsolanki
Messages
77
Reaction score
0
I'd like to start off by noting that cellphone billing is very different from that in the US. The consumer is free to choose any cellphone he wants and he still has a choice for the carrier/telecom service provider. This is the reason why unlike in the US, providers in India depend on earning the patronage of particular demographics through customized plans for them. DIfferent plans for collegians, for people who want just incoming calls, people who have to make a lot of short calls etc.

- - - - - - - - - - - - - - - - -

I am quoting the problem statement verbatim - the next post will elucidate the solutions that I am working on.

The problem is to analyze the past behavior of the subscribers and build a model
which recommends the optimal plan for them. The past behavior of the subscriber is
recorded in the form of call detail records (CDR). CDRs are generated for every call the
subscriber makes in the network. Typical information recorded as part of the CDR
includes Calling Number, Called Number, Current Plan, Service (Voice/SMS/etc.),
Duration, Cost, and Location.
The model should address the following:

1. Changing behavior of the subscriber
2. Should be parallelizable (A parallel model allows an alternative representation of the
state of an application, either at a different time or in a hypothetical state).

The primary question here is : What branch of mathematics and WHAT topic should I refer so that I can attempt this question methodically?
 
on Phys.org


We have till now classified the various types of people/characters who use a phone and listed out the particular traits of their usage. Now we will think backwards and analyse their usage by studying their CDRs. We want to know how we can draw inferences from CDRs.

Inferring data usage from averages might be misleading so we can't just define a few bill plans and charge the customer by the plan by which he gets minimum bill for the last month. So we need a way of selecting a plan using their previous usage pattern but not by taking averages. So average can't be the tool for inference. The question is: What can be?
 


Sounds like you'd like to find appropriate summary statistics (also called descriptive statistics) for the frequency/duration/cost data. Examples of summary statistics include average, minimum, maximum, quartiles, variance, etc and there are plenty of good articles on the web.

Good luck with it!
 


Anything on clustering?
 


I would also suggest mathematical optimization, linear or nonlinear.