Kasper Van Lombeek is a passionate statistician. After working very hands-on in the offshore wind industry, he realized he was stuck with his Excel files and went to study statistics. Over the last few years he has been focussing on lots of other aspects regarding data science, such as building interactive applications. He co-founded the company Rockestate, which applies AI on open geo data. The combination of open geo data with internal data of a bank, insurance or energy company brings radical change in the way risk is managed and products are priced.
Case Study: Recommendation Engines: optimizing mean square error or user experience?
Since the 1 million euro contest to improve the performance of the Netflix recommendation engine, recommender algorithms are one of the hottest topics in data science today. That is why we, passionate data scientists, had to build our own recommendation engine. As young fathers with lots of friends having babies, we asked ourselves: can we build a service that finds out your taste and recommends a baby name for you? Today we have a solid website www.namesilike.com that does exactly that.
With today’s online tutorials, it is not very difficult to learn how to build a recommender system based on a dataset. Those tutorials only talk about mathematically optimizing the algorithm. But that is not what you have to optimize! You have to set up a whole process online, starting from finding out the user’s taste with the least possible effort, and ending with showing the recommendations. The difference between optimizing the accuracy versus the user experience is enormous. For example: some of the most interesting algorithms create beautiful visualizations of your products. Although you lose a bit in prediction power, the gain in user experience is enormous. In our talk we highlight some of these trade-offs. They are all real lessons learned on Names I Like, validated by thousands of user clicks.