Making the Leap in Machine Learning

April 24, 2019

Kount is advancing supervised machine learning model execution with MLeap. Find out what this fraud prevention advancement means for our customers.

A little background

Noah Pritikin, Lead Site Reliability Engineer for Kount’s AI Team, investigates enhancements to Kount’s products and technologies. “We were researching ways to move the needle in Boost, one of our supervised machine learning* products,” he says. In conjunction with Kount’s AI team, he discovered a method to deliver fraud prevention predictions to our customers faster than it’s ever been done before at Kount. By connecting the right platforms with Kount’s capabilities, the team increased prediction speeds by hosting supervised machine learning models with MLeap. Noah describes the team’s ingenious deployment.

Machine learning is one important component of Kount’s fraud protection methodologies. Based on algorithms and statistical models, systems can learn from data in order to identify patterns and make recommendations with minimal human involvement. It’s no small feat to reduce the average response time to generate transactional predictions by more than half (from 19.27ms to 7.00ms) while increasing the scalability and reliability of available data. That’s the innovative breakthrough Noah and team accomplished.

Digging deeper into data processing

Since introducing BOOST Technology* in October 2017, Kount continues to advance technologies to enhance the BOOST supervised machine learning model with tools such as Apache Spark. One of the most adopted big data distributed processing frameworks in the world, Apache Spark is an open source, general-purpose distributed computing engine used for analyzing large amounts of data. Fast, flexible, and developer-friendly, Apache Spark is the leading engine for large-scale batch processing, stream processing, and machine learning.

Its primary advantage is the ability to parallelize work — decreasing the time to get your answers. Spark supports in-memory processing to improve the performance of big data analytics. This means that it can perform tasks up to one hundred times faster than other engines in certain situations. Application developers and data scientists can easily harness its scalability and speed to compute data across industries. 

It was a hunch that proved to be the right one

Prior to using Spark to generate the Boost supervised machine learning model, Kount used Python3 with Scikit-learn. This worked well as a first-generation tool, however it posed limitations, including lack of portability, inability to scale, lack of multiple model support, and limited model governance.

In a continuous effort to push the envelope, the team faced the challenge of bringing a Spark-generated machine learning model into the production environment. Through research around how to “productionize” Spark-generated machine learning models, the team discovered MLeap and a proof of concept was born.

MLeap enabled Noah and team to optimize the processing speed of a high volume of data with minimal delay. “By applying well-known, open source capabilities to Kount’s platform, we achieved the low latency results that are required to make machine learning model predictions in a production environment a reality,” Noah says. “We have more capacity to process fraud at a higher rate,” he adds.

The response time statistics for generating transactional predictions were impressive. Before utilizing MLeap, the following defines Kount’s latency characteristics:

Average 95th Percentile 99th Percentile Standard Deviation
19.27ms 24ms 37ms 5.31ms

After including MLeap in our “real-time” machine learning execution architecture, the latency characteristics are the following:

Average 95th Percentile 99th Percentile Standard Deviation
7ms 9ms 16ms 2.41ms

The results are impressive: Roughly 64% faster on average, and 57% faster across 99% of all prediction requests.

What’s the impact?

By enhancing its machine learning characteristics, Kount adds greater capacity to process potential fraud at advanced rates. “We have created new value for our products,” Noah says.

Kount’s Boost Technology takes advantage of the billions of data points available through its global network. With this new solution, Kount continues to move the industry forward with speed and innovation. 

Would you like to learn more? Read this technical white paper and/or contact Noah Pritikin.

* Noah Pritikin is the lead SRE for Kount’s AI Science Team. His broad, multi-discipline approach provides the AI Science team with innovative, scalable, and robust solutions to fight against digital fraud. These solutions operate in less than 250ms to protect hundreds of millions of dollars of payment transactions each day. Previously, at The Boeing Company, he led a team within the Global Network Engineering organization that increased first time quality and saved tens of thousands of dollars in labor through the creation of many automated solutions. He has a BS in Computer Engineering from California Polytechnic State University in San Luis Obispo, CA.

* Machine Learning is the scientific study of algorithms and statistical models that computer systems use to progressively improve their performance on specific tasks. Machine learning algorithms build a mathematical model of sample data, known as “training data,” in order to make predictions or decisions without being explicitly programmed to perform the task.

* Boost Technology aggregates millions of transactions and their outcomes, including approvals, chargebacks, refunds, and reviews. Boost Technology is designed specifically to weigh the risk of fraud against the value of each unique customer and help to identify legitimate transactions from fraudulent ones. The machine learning featured at the core of Boost Technology works out-of-the-box, evolving and improving its algorithm as it builds a unique picture of each merchant’s business.