Group Abstract

Message Boards

WOLFRAM COMMUNITY

18.2K Views

11 Replies

3 Total Likes

View groups...

Follow this post

Share this post:

GROUPS:

Find the optimal virtual phone numbers to buy for X customers area codes?

David Johnston

David Johnston, Artificial Brilliance, Inc.

Posted 8 years ago

We have a unique problem that not even large telecom carriers have solved yet. I am hoping I could get some insights from the community on potential solutions since it is probably a weighted distribution problem with matrix multiplication or something else. Either way, I think it is a math question and there is no better place to look for math related programming solutions than here. :) Thank you ahead of time! We need a mapping algorithm that takes X number (lets say 500,000) of Customer Contact Records and figures out the best X number (lets say 1,000) of Virtual Phone Numbers to buy and assign to the contacts. The assignment follows this order of priority: area code, zip code, city, state and country. The customer contacts have numbers from all types of area codes. Mainly the quantities follow city population for the cities in or close to the area codes. So, of the 500k contacts, usually no single area code has more than 10% representation. Normally its only about 1-4% of any customer list is in a single area code. So, the challenge is to create an algorithm that looks at the frequency of area codes in the 500k list and chooses the 1k best area codes to buy virtual numbers in. Here is where it gets hairy. No virtual number can be assigned to more than 250 contacts in the same area code. So, the algo needs to tell us how many of each area code we need to buy numbers in. Area code: 951 may have 600 contacts in it. In that case, we need to buy 3 numbers in that area code. That means 3 of the 1,000 are a single area code. That type of thing has to be calculated in the 1k virtual number buy recommendation algo. If there is no geo match possible and all contacts are in the same country, the contact is assigned to a random virtual number. So, we need a buy algo and a mapping algo. To make things more complicated, if there is an existing mapping but user buys more numbers, we need to allow the user to remap all contact to virtual number relationships in a way that maps the closest geo relationships. As always, if you need a bounty to provide me a solution, please let me know. I have always paid immediately in the past.

POSTED BY: David Johnston

11 Replies

Sort By:

CJ B

Posted 8 years ago

Is this for optimizing "RoboCalling"? If so, I would propose that the thread be deleted from the forum by the moderators because while RoboCalling is legal, it probably should not be legal (and certainly should not be sanctioned by Wolfram or anyone else).

POSTED BY: CJ B

Sander Huisman

Sander Huisman, University of Twente

Posted 8 years ago

I think I understood the problem a bit better now, though not 100% clear yet. Imagine you have all the lat, lon coordinates (geo-positions) of your clients, whatever the amount. What I would do is: Create n number of random points inside the bounding box of these points. Where n is the number of virtual phone numbers. Calculate the Voronoi Tessellation of these points Perform Lloyd relaxation on this tessellation a few (10?) times to make the points uniform. Calculate the number of customers that fall inside each Voronoi cell. This will be the 'counts'. Now again, create random points but increase the chance of a point being somewhere depending on the number of counts. Perform 1 or 2 cycles of Lloyd relaxation to smooth it out a bit, and count again, Now the counts of each polygon should become similar. Repeat 3-5 as many times necessary... Now you have some Voronoi cells left for which the counts should be roughly constant. i.e. each region for the phone-numbers is convex and local (what is wanted/required I presume) and each region has roughly the same number of customers.

POSTED BY: Sander Huisman

David Johnston

David Johnston, Artificial Brilliance, Inc.

Posted 8 years ago

Also, this provides a huge benefit when it comes to analytics. We can now break down by geo as to what is working or not when it comes to both their advertising but customer service as well. So, imagine there are 50 virtual numbers in the 951 area code. If each of those are grouped by zip first and city second, you can see how quickly the inbound/outbound call and text stats begin to provide a massive edge to the enterprise. They can then narrow down and say x number of phone sales came from x location. They can also analyze angry customer service calls and pin point an approximate geo-location and take appropriate action to replace or retrain the manager in the geo. They can additionally do geo-sentiment analysis on both inbound text messages as well as phone logs. It's not a new idea but this type of segmentation technique has been coined "hyper-geo" and for good reason. You may think "they could run these same analytics based on the address they have on file for the customer." This is true in some cases but many times the address on file is not accurate. Additionally, it also requires the enterprise to cross reference databases, etc. and it becomes an IT nightmare. We are skipping IT and providing hyper-geo analytics to their marketing and sales department without them having to get their IT involved. Additionally, our carriers are going to start providing us with "Last Known Triangulation" of the cell phone. We will add that as an additional column and in priority order it will be before zip code. It will be lat and long and that is it. Some say Google analytics or other products provide great geo-analytics. However, they are not monitoring billboards and other signage, etc etc.. They are also not showing the true keywords the user typed in and it shows in the dashboard as (Not Set). No analytic platform that I have tried and tested works for massive hyper geo needs. With this system, they can use a different number for every ad and have cross platform and cross device tracking because their tracking is not based on the type of advertising they are showing people. Here is an example of Hyper GEO Google Pay Per Click Software that we built. After watching these, you can see why hyper geo phone numbers are important for advertising. I think the customer service use case is a little more obvious. 1 Video: https://www.youtube.com/watch?v=kSN5lWbwWzI 2 Video: https://www.youtube.com/watch?v=QAVf5CI3vIs

POSTED BY: David Johnston

David Johnston

David Johnston, Artificial Brilliance, Inc.

Posted 8 years ago

Geo-distribution is best because a single virtual number is not recommended to have more than 250 contacts assigned to each. So one number that covers tons of people is actually counter productive. That is why we need one algo for generating the recommended buy list and a second for dealing with the mess of remapping based on closest geo.

POSTED BY: David Johnston

Neil Singer

Neil Singer, AC Kinetics, Inc.

Posted 8 years ago

What makes a particular "number buy" best? One that covers the most number of the 500k people? One that has the most geographical distribution?

POSTED BY: Neil Singer

David Cardinal

David Cardinal, Cardinal Photo

Posted 8 years ago

As a start, maybe K-means would get you a rough solution (assuming you can set the parameters to match your proximity objectives), although without some of the constraints you mention.

POSTED BY: David Cardinal

Sander Huisman

Sander Huisman, University of Twente

Posted 8 years ago

I think this question becomes much more clear if you provide a small (bogus) dataset and the expected outcome (with a cap of 5 instead of 250). Especially the priority part is not that clear...

POSTED BY: Sander Huisman

David Johnston

David Johnston, Artificial Brilliance, Inc.

Posted 8 years ago

Yes. Large enterprise customer needs to be able to spread out their voice and sms traffic over a large pool of numbers. However, it is not good for their relationship with their customers to not provide a local number they can call and text to. The closer the geo-match, the more their customer engages with them. That equals better customer service experiences. Despite what some telecoms may claim, it is impossible to handle millions of inbound calls and texts to a single phone number simultaneously. It is better to distribute the traffic over a pool of numbers. If that pool contained geo-clusters, that would be ideal.

POSTED BY: David Johnston

Christopher Wolfram

Christopher Wolfram, WOLFRAM

Posted 8 years ago

I'm not sure I entirely understand the question either, but it seems that the critical point is that you want to assign customers someone with the same area code so that they can make local calls. But does a smaller GeoDistance really matter, or is it just whether it is a local area code? If the virtual phone has an area code for the neighboring town, is there any advantage or is it just as good as having an area code from across the country?