How to train your model (and catch label leakage)

Description

A pervasive but often overlooked problem in predictive modeling on real-life data is the problem of data or label leakage. At enterprise companies that provide ML as a service to other businesses, such as Salesforce, this problem takes on monstrous proportions as the data is populated by diverse and often unknown business processes, making it very hard for data scientists to distinguish cause from effect.

Till Bergmann explains how Salesforce—which needs to churn out thousands of customer-specific models for any given use case—tackled this problem. The automated approaches are a part of our recently open-sourced Spark-based library TransmogrifAI and extend the boundaries of what typically falls in the domain of automated machine learning.