apply() - Training, Hosting and Deploying 1 Million+ Personalized Models

Speaker: Ian Sebanja and Simarpal Khaira, Intuit

"Quickbooks online(QBO) is one of the core Intuit products that leverages ML to provide an awesome experience for users. Customers come to QBO for business insights, but often find themselves swamped in the work of categorizing a multitude of banking transactions. Traditionally, the ML model used to categorize transactions for Quickbooks online users was a general model that was applied to all businesses irrespective of industry. While the results were good, there was room for improvement.

The solution: Intuit built a categorization model suite, including (1) a global model which leverages populational data to help users to categorize their transactions and (2) personalized models to adapt category recommendations to each company's chart of account. This involved creating 1.8 million unique models. One user one model, trained on their features, hosted and produced predictions for a specific user and refreshed based users interaction signals. Model management for 1.8 million models required the Machine Learning platform to come up with solutions for featurizing, training, hosting(deploying) and monitoring these models. This talk will focus on our approach to solving this problem for our platform users."