Prediction of bike rental using model reuse strategy

This paper describes the methodology used for ECMLPKDD 2015 Discovery Challenge on Model Reuse with Bike Rental Station Data (MoReBikeS). The challenge was to predict the number of bikes in the new stations three hours in advance. Initially, the data for the first 25 new stations (station 201 to 225) was provided and various prediction methods were utilized on these test stations and the results were updated every week. Then the full test data for the remaining 50 stations (station 226 to 275) was given and the prediction was made using the best method obtained from the small test challenge. Several methods like Ordinary Least Squares, Poisson Regression, and Zero Inflated Poisson Regression were tried. But reusing the linear models learnt from the old stations (station 1 to 200) with lowest mean absolute error proved to be the simple and effective solution.