Ideas

[ankane]

Ideas

• Add support for getting probabilities for classification
• Add random hyperparameter search - https://lightgbm.readthedocs.io/en/latest/Parameters-Tuning.html
  • try 30-100 combinations
  • learning rate (10^-4 to 1, log scale)
  • depth (2-6)
  • feature_fraction (.5 to 1)
• Drop support for gsl gem (no longer maintained and Eps supports gslr gem)
• (maybe) Move to ONNX format

[HansHauge]

Hey @ankane, is there currently any way to change the learning rate when creating a new model?

When I try model = Eps::Model.new(data, target: :price_usd, split: :unix_timestamp, learning_rate: 0.01) for example, I get ArgumentError: unknown keyword: learning_rate.

Is this a currently supported feature and I just need to enter the argument differently, or is this feature currently not supported?

Thank you.

[ankane]

Hey @HansHauge, added on master.

[HansHauge]

Thank you, this is a great addition. That one change makes the LightGBM much more powerful for the user!