Hacker News new | ask | show | jobs
by moultano 3326 days ago
Much better to include the feature in the training, and exclude it from the inference. That actually cancels out the effect, rather than just incentivizing the model to reconstruct it.