Predictive Analytics or Data Science?

Monday, February 9, 2015 - 19:45

I caught up with an old grad school friend a few weeks back. He's a top-notch statistician who's built a successful career working in quants departments of large insurance and health care companies. With little simplification, I'd characterize his role over the last 20 years as a predictive modeling expert. His work is primarily “big iron” -- revolving on Teradata, Oracle and SAS. Besides being a senior statistician, he's also a more-than-capable data integration and statistical programmer.

In the past few years especially, we've had “discussions” on the differences between data science (DS) and statistics/machine learning as disciplines. He's characterized DS as little more than a trumped up moniker marketed by the newest analytics generation to brand themselves with a sexy statistics job title – for work that's indistinguishable from what he's been doing for years.
I've disagreed, not only seeing DS in the service of new data/analytics products, but also arguing the big data, integration, and computation obsessions of data science differentiate it from traditional statistics. More and more, it's data/computation at the forefront, with the likes of Hadoop/Hive/Pig, Spark, R, and Python/Pandas/scikit supplanting SAS in the data warehouse as the go-to tools for DS types. And yes there's a generational divide in usage of the former and the latter. It's millennials vs baby boomers.

This time my friend took a new tack, challenging me to distinguish not statistics from data science, but predictive analytics (PA) from data science. And, I must admit, that gave me pause.

PA & DS both contrast with statistics in their emphasis on prediction over causality and their general use of observational in contrast to experimental methods. In addition, I've always seen predictive analytics as applied statistics/machine learning in the work world, more data-focused and computational than statistics, but less so than data science. When challenged to define the ”point” that separates PA from DS, however, I couldn't, arguing feebly there's a continuum from statistics to data science on a data/computation axis with endpoints “not so much” and “lots” -- and predictive analytics in the middle. That's certainly how my company, Inquidia Consulting, sees it.

It was only after we parted that it occurred to me the differences between PA and DS were clearly manifest in two current Inquidia projects.

The predictive analytics engagement first articulated a classification prediction challenge, after which roughly identifying the source of features from the data warehouse. Early on, there was lots of data extract and munging with SQL and Pentaho, followed by exploratory data analysis and machine learning algorithms in R. In the end, the deliverable was a cross-validated and surprisingly accurate Multivariate Adaptive Regressive Splines model with in-the-future prediction capability. Analytical effort allocation: 25% data/computation, 75% stats/ML.

The data science effort was driven by an association challenge not amenable to collaborative filtering or other off-the-shelf algorithm. Instead, the association was split into a large number of binary classification tasks, which could then be subjected to wide variety of dichotomizers. Indeed, the data/computation work took several mathematical and engineering optimization turns. Ultimately, an appropriate classifier was chosen, a means for training devised, and simplifying assumptions that allowed individual decisions to select the most likely association divined. The prototype was developed in Python, using Apache Spark, Pandas, NumPy, and SciPy, sourcing data from Hadoop/Hive. In the end, a modestly successful model plagued by sparse data was delivered. The production implementation deployment is being written in a combination of Hive and Pig, with Java user-defined functions. Analytical effort allocation: 75% math/data/computation, 25% stats/ML.

Not contented with this soft assessment, I set out to “validate” the observation from a series of analytics LinkedIn job postings. I found eight ads with “data science” in the title, then separated the data/computation verbage from statistics/machine learning. The table below details the cross-tabulation.

The admittedly limited and likely biased sample suggests that data/computation is at least the equal of statistics/ML as a must-have for data science wannabees. Indeed, the data/computation reqs seem more focused than the stats/ML, intimating they have primacy in the early winnowing. It'd be interesting to do a larger examination of characteristics of winning candidates to test if the data/computation “hypothesis” holds up.

Contact us today to find out how Inquidia can show you how to collect, integrate and enrich your data. We do data. You can, too.

Would you like to know more?

Sign up for our fascinating (albeit infrequent) emails. Get the latest news, tips, tricks and other cool info from Inquidia.