Daniel Haugstvedt

Because ML research focuses on models, it is easy for industry practitioners to fall into the trap and do the same. The use of Kaggle as a platform for learning ML amplifies this problem. Ruter’s machine learning team spends less than 5% of their time on model development.

Where do they spend their remaining time? They are scoping projects, identifying use cases, and generating value from existing models. Because of this, Ruter has flipped the deck and switched from developing machine learning models based on performance metrics to developing models based on usability.

Here we will present the evolution of our load prediction model. It predicts the load for ~500 000 daily departures across ~200 routes for busses, trams, and boats with more than ten different types of sensors.