Jaap Suermondt (HP Labs) talked about how his lab uses analytics to support operations at HP. Jaap started with an example of procurement of disk drives. HP ships two PCs per second and buys more disk drives than anyone else so they have to have accurate estimates of demand. This is roughly equivalent to predicting how the economy and stock market will do. But by combining genetic algorithms and economic and statistical models, HP was able to predict that the economy would improve and demand would improve by five percentage points more than others predicted. This was the subject of a Business Week article and video in 2009 that talked about how earlier genetic algorithms work was revived and applied to this problem.
HP also forecasts demand for labor so as to be able to avoid having to fire or hire people. When people are overcommitted, attrition goes through the roof. When people are idle, production costs are unnecessarily high and this is an issue in PC production because it is so competitive and the margins are low.
Another example involved maximizing the revenue from covered orders (RCO – Revenue Coverage Optimization). Trying to maximize revenue by individual products produces results worse than random. It’s important to consider combinations because of dependencies between products (e.g., it may not be possible to assemble a highly profitable order for a customer because some relatively unprofitable item is unavailable). When HP ranked their products by importance to revenue coverage, they got a nice Pareto effect: 80% of the revenue with 25% of the products. This work won the coveted INFORMS Edelman prize in 2009.
Interestingly, this example runs counter to the idea that a lot of data makes it a bad idea or unnecessary to improve one’s algorithm. There was no shortage of data in this case and in fact there was so much data that the initial algorithm, an integer programming algorithm, was too slow to be useful. Reformulating the problem as a Lagrangian Relaxation problem shortened the solution time. But the breakthrough in this case involved reformulating the problem as a bipartite graph flow problem and coming up with a new algorithm for solving this problem using the data in real time.
In addition to operations, Jaap’s lab covers collaborative filtering, customer segmentation, marketing analytics, personalization, recommendation, and so on. He views personalization as the killer app for consumer facing services. He argued that it is crucial to lead with the customer experience and make it a win-win so that companies don’t face angry customers and backlash against privacy violations.
Another non-operations example Jaap described involved analytics work with Stanford Children’s hospital that has saved over thirty children’s lives. They analyzed patient’s data to find indicators that children were in trouble and used them to trigger more rapid responses. There is a huge number of preventable deaths (100,000!) in the United States each year so there is a lot of room for more work using analytics to help save lives.
The video for Jaap’s talk is available at dyyno.com.