PAPIs 2018

Machine Learning applications at Salesforce use a wide variety of customer data that is highly customizable. In this talk, I discuss some challenges of designing automated machine learning pipelines that can deal with custom user data that it has never seen before, as well as some of our solutions. Examples include statistical tests between training and scoring data sets to help with the cold start problem, algorithms to throw out features that are "too good" because they are derived from the label we're trying to predict, and data-dependent feature engineering steps like automatically determining buckets for numeric variables and detecting categorical variables encoded as other data types.