Building Data Science at Scale
As part of the Boston-based Engineering group, the Data Science team’s charter is to enable science-based personalization and recommendation for PayPal’s global users. As companies of all sizes are starting to leverage their data assets, data science has become indispensable in creating relevant user experience. Helping fulfill PayPal’s mission to build the Web’s most convenient payment solution, the team works with various internal partners and strives to deliver best-in-class data science.
Technology Overview
At the backend of the data science platform reside large-scale machine learning engines that continuously learn and predict, from transactional, behavioral, and other datasets. An example of a question we might try to answer is: if someone just purchased a piece of software, does it increase his likelihood to purchase electronics in near future? It is no wonder that the answer lies in the huge amounts of transaction data, in that what people bought in the past is predictive of what they might consider buying next.
We leverage state-of-the-art machine learning technologies to make such predictions. Machine learning itself has been quickly evolving in recent years. New advances including large-scale matrix factorization, probabilistic behavioral models, and deep learning are no strangers to the team. To make things work at large scale, we leverage Apache Hadoop and its rich ecosystem of tools to process large amounts of data and build data pipelines that are part of the data science platform.
Tackling Data Science
Companies take different approaches in tackling data science. While some companies define a data scientist as someone who performs statistical modeling, we at PayPal Engineering have chosen to take a combined science & engineering approach. Our data scientists are “analytically-minded, statistically and mathematically sophisticated data engineers” [1]. There are of course more science-inclined, and more engineering-inclined individuals on the team, but there is much more of a blend of expertise than a marked distinction between these individuals. This approach to data science allows us to quickly iterate and operationalize high-performing predictive models at scale.
The Venn diagram below, which bears similarity to Conway’s diagram [2], displays the three cornerstones pivotal to the success of the team. Science, which entails machine learning, statistics and analytics, is the methodology by which we generate actionable predictions and insights from very-large datasets. The Engineering component, which includes Apache Hadoop and Spark, makes it possible for us to do science and analytics at scale and deliver results with quick turnaround. Last but not least, I cannot emphasize more the importance of understanding the business for a data scientist. None of the best work in this area that I know of is done in isolation. It is through understanding the problem domain that a data scientist may come up with a better predictive model among other results.
Scaling Data Science
There are multiple dimensions to scaling data science, which at a minimum involves the team, technology infrastructure, and the operating process. While each of these is worth thoughtful discussion in its own right, I will focus on the operating process since it is critical to how value is delivered. A typical process for the team could break down into these steps:
- Identify business use case;
- Understand business logic and KPIs;
- Identify and capture datasets;
- Proof of concept and back-test;
- Operationalize predictive models;
- Measure lift, optimize and iterate.
Let me explain some of the key elements. First, we see scaling data science as an ongoing collaboration with Product and Business teams. Understanding product and business KPIs and using data science to optimize the same is an essential ingredient of our day-to-day. Second, we follow the best practice in data science, whereby each predictive model is fully back-tested before operationalization. This practice guarantees the effectiveness of the data science platform. Third, the most powerful predictive modeling often requires iterative measurement and optimization. As a concrete example, putting this process into practice along with PayPal Media Network, we were able to achieve excellent results based on:
- Lookalike modeling: Merchants can reach consumers who look like the merchant’s best existing customers.
- Purchase intent modeling: Merchants can engage consumers who have a propensity to spend within specific categories.
While it is challenging to crunch the data and create tangible value, it is interesting and rewarding work. I hope to discuss more details of all the fun things we do as data scientists in the future.
References:
[2] http://drewconway.com/zia/2013/3/26/the-data-science-venn-diagram
