The goal of this problem was to replicate results obtained by Ehsan Imani in his thesis and paper about the Histogram Loss. It contains old experiment code by Ehsan Imani that has been updated to Python 3. Additionally, it contains new code for running experiments on these datasets with our histogram loss framework to compare to the original results.
- CT Position (53500, 386) - Predicting the relative location of a CT slice within the body (Graf et al. 2011)
- Bike Sharing (17389, 16) - Predicting the hourly count of rented bikes (Fanaee-T and Gama 2014)
- Song Year (515345, 90) - Predicting the release year of songs from audio features (Bertin-Mahieux et al. 2011)
- Pole (15000, 49) - Describes a telecommunications problem (Olson et al. 2017)
The original experiment used MLPs with 4 hidden layers and dropout of 0.05 on the input. The hidden layer sizes are half of the input size, except for bike sharing which uses 64.