The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Deep Learning with Structured Data w/ Mark Ryan - #301
Sep 19, 2019
Mark Ryan, author of 'Deep Learning with Structured Data' and a member of IBM's Data and AI support team, shares insights on applying deep learning to structured data. He discusses creating predictive models using the Toronto streetcar network dataset, addressing challenges like data preparation and metadata integration. Mark emphasizes that deep learning doesn't always require massive datasets and highlights its potential in various sectors. He also details the interactive feedback process in his book's development and the advantages of collaborative learning in deep learning courses.
AI Snips
Chapters
Books
Transcript
Episode notes
Mark's Deep Learning Journey
- Mark Ryan's interest in deep learning was reignited by its resurgence around 2016, leading him to explore its applications in structured data.
- He developed prototypes for predicting ticket closure times and duty manager calls at IBM using proprietary data, before seeking a public dataset for broader application and sharing.
Trouble Ticket Textual Data
- Mark used both structured data and textual descriptions from trouble tickets to predict close times.
- Including the descriptions, processed with an RNN, improved accuracy by 3-4%, highlighting the value of combining data types.
Textual Data Impact
- Adding descriptive text to structured ticket data yielded only a 3% accuracy increase in close time prediction.
- This suggests the structured data itself holds significant predictive power, diminishing the relative impact of textual data.
