"Scaling and Approximation in Complex Data Analysis", Mikio Braun

Scaling and Approximation

in Complex Data Analysis

Mikio BraunDelivery Lead Recommendation & Search

Zalando SEdatanatives, Nov 20, 2016

● 135M visitors per months (about 4.5M per day)

● Recommendations for product page, newsletter, etc.

● Fully data driven.

10.04.15

Recommendations at Zalando

Personalized Recommendations

Big data on Hadoop

Large Scale Learning

Learning in a nutshell: Optimization

Instead of doing anything “fancy”,try to minimize the prediction error in each step.

Or, not even consider the whole data set but just a few points at a time.

Or, not even that but only take one point at a time.

=> Stochastic Gradient Descent

Gradient Descent - Introducing approximations

Why can we just take a short-cut?

● Large scale complex data analysis: billions of examples, millions of features

● Many parts can be parallelized well● Training of models is essentially hard● Approximation can help to deal● Goal is to generate good predictions of future data

Summary