Scoring 160 million companies for propensity to buy at Cisco

"Decision trees are good. They're quick. They're reliable, but not the best in terms of generating predictions," said Lou Carvalheira, Advanced Analytics Manager at Cisco. With a 20-person team scoring a database of 160 million companies, the factory needed to run faster and reach the full dataset.

In 2014 Cisco adopted H2O's open-source, in-memory distributed machine learning platform and rebuilt the factory around it, with the whole process flow controlled in R. Ready-to-use algorithms let the team move past decision trees to GLMs, random forests, GBMs, and deep learning: a first-stage model estimates the probability of purchase and a second stage the expected value, then the best model, or an ensemble, does the scoring.

H2O's in-memory engine cut the need for expensive storage and let Cisco train on the entire dataset rather than a sample. On a four-node cluster with 24 cores and 128GB of memory each, Cisco now scores all 160 million companies in a couple of days, and new buying data flows into the models immediately instead of being prepared in advance.

What once took more than a month now takes two days, a 15x increase in speed, on a dramatically larger dataset. Cisco deploys roughly 60,000 propensity-to-buy models every quarter across hundreds of millions of observations, using deep learning and GBM rather than simple decision trees for more accurate predictions, and folding in new buying data with no advance preparation.

Since its inception the factory has been associated with more than $3 billion in revenue. "We use H2O where it shines in terms of model training and scoring," said Carvalheira.