Thank you so much for your thoughtful response. We appreciate the gracious response given that the post highlighted some concerns with the original work.
It is somewhat challenging to create principled methods for selecting the hyperparameters you mentioned, and this seems like an area for future work. We also experimented with hyperparameters such as λ = −0.25 which produced an even higher R^2 to the open source models of 0.925, but which produced predictions of the proprietary models which seemed quite low to us.
The method certainly seems promising and we are excited to see your future results!
Thank you so much for your thoughtful response. We appreciate the gracious response given that the post highlighted some concerns with the original work.
It is somewhat challenging to create principled methods for selecting the hyperparameters you mentioned, and this seems like an area for future work. We also experimented with hyperparameters such as λ = −0.25 which produced an even higher R^2 to the open source models of 0.925, but which produced predictions of the proprietary models which seemed quite low to us.
The method certainly seems promising and we are excited to see your future results!