Product Update: Name Ranking for Reverse Phone API & Caller ID API

November 16, 2022
Product & Engineering

One of the areas of focus in the last couple of months at Trestle is accuracy. We measure that by checking if we have the right name and address associated with the phone number. While our accuracy numbers have been great as mentioned in this post, one of the focus areas has been how we rank the names/addresses associated with the phone. Meaning, if we have multiple people, we should rank the most likely person using the phone as the highest. 

Here is an example of what I mean. Our database has multiple names associated with this phone. This can happen because there is a single account holder registered with the Telco, but different people in the household are using the phone.

Ned Stark

    • Phone: 555-555-1212
    • Current Address: Winterfell, The North
    • Historical Address: House of Lannisters
    • Household People: Robb Stark, Sansa Stark

Robb Stark

    • Phone: 555-555-2424, 555-555-1212 
    • Current Address: Winterfell, The North
    • Historical Address: House of Lannisters
    • Household People: Ned Stark, Sansa Stark

Sansa Stark

    • Phone: 555-555-3636, 555-555-1212
    • Current Address: House of Lannisters
    • Historical Address: Winterfell, The North
    • Historical Address: Household People: Robb Stark, Ned Stark

In this case, it is clear that the 555-555-1212 number belongs to Ned Stark, but probably the other 555-555-3636 number is a better match for Sansa Stark (rather than Ned). Hence, when a call comes in, businesses will be better informed and can take appropriate action on lead prioritization or for call routing purposes if they know that Sansa is using the phone from the House of Lannisters.  

This is a challenging problem to solve as data providers like Telco do not have this fine-grained data and the ranking challenge. Fortunately, our data partners, as well as our own internal data, have quite a few signals that can help us infer the ‘best match’. That’s exactly what we have been iterating over the last couple of months, and we are super excited to bring it to our customers. Let’s just use the numbers from our truth set which will explain our excitement. 

Before:

    • Our accuracy where we had a name associated with the phone: 93%
    • Our accuracy with the rank-1 name as being associated with the phone: 76%

Now:

    • Our accuracy where we had a name associated with the phone: 93%
    • Our accuracy with the rank-1 name as being associated with the phone: 87%

Again, there’s more work to do, but there’s one thing I will emphasize:  With any change, we do truth set analysis to measure our improvements and to see how we are doing against our competitors. There is literally no provider out there with this kind of accuracy at this coverage level. Period. We have had customers already tell us about our coverage and accuracy, but this change now puts us miles ahead of the competition. 

 

A few challenges we had to overcome in order to deliver this update:

    • Building a big enough truth set so that we can understand the ranking patterns based on the 100+ signals we look at.
    • Figuring out the signals of interest and which make the most sense in the model. This is as much art as it is a science, and only comes from a lot of learning over the years on what adds value and what is noise. Finding these signals from everything we get from our data partners, cleaning it up and then feeding into the model was probably the 80% work here as any data scientist will attest to.
    • A machine learning model that can ingest these signals and determine the ranking in real time. While we have machine learning models already for merging the data from various providers, this one determining the rankings is a different model, and with any new model, it took quite a bit of undertaking.
    • As alluded above, we need to make this ranking determination in real time as some of the signals are received in real time from our providers. The work we had to ensure this ranking with minimal latency was another huge engineering accomplishment.

 

All of this is worth the effort given the level of improvements we have been able to make. We are looking forward to seeing this improvement make a difference for our customers in their conversions. More to come of course, and pretty soon so stay tuned!

If you have any questions, need any assistance with our products or just want go through our truth set analysis, you can reach me at kshah [at] trestleiq [dot] com.

Developer Portal