Tom Tom Festival 2019

This session is meant to summarize hard lessons learned from decades of data processing, data analysis, data mining, and recently data science.  We discuss the subtleties of GIGO -- Garbage In, Garbage Out -- and how this lurks within every data task we undertake, from deep learning to service orchestration, in ways that can be difficult to recognize.  Real-life case studies highlight the various ways in which well-meaning data scientists and their clients can do harm if we do not practice our core skepticism.  These personal lessons learned and supporting examples are organized into three themes:

• the data is not the thing
  
  • a medical model to bolster physician exposure to rare cases:  mistaking choices we made for choices we should have made
    
• the model is not the thing
  
  • predicting fraud to generalize a curated set of rules:  mistaking machine learning for human learning
    
• the data scientist is not the thing
  
  • a financial-services model of customer satisfaction:  mistaking modeller integrity for solution integrity
    

The synthesis / take-home message -- what the thing IS -- is to recognize modeling as metaphor:  Every instance is imperfect, but some can be useful.  The right approach is to iterate on not only modeling and development but on communication and cross-expertise vetting so as to constrain and characterize the imperfections.


You need this ticket from Eventbrite to sign up: Applied Machine Learning Conference.