Predicting Passenger Counts in Rail Traffic – Heikki Pulkkinen, VR Group

Using XGBoost to predict the number of passengers in each train at a given point in time. One of the most important parts of our architecture is measuring the training-serving skew from the multiple different models in production.


  • Getting something into production is more important than fine-tuning the model
  • A big regression model can combine multiple classical approaches in a easily managed way
  • Having the right tools (databases and processing environments in cloud) make the way much easier

Add comment