negative 'Start training from score'

user_13508767 0
Viewed 314

When running lgb.cv, from the log I sometimes get negative numbers following 'Start training from score'. Wondering what does the number actually mean, and what unit? Is it in terms of the metric specified in params? Here's an excerpt:

[LightGBM] [Info] Total Bins 2767
[LightGBM] [Info] Number of data: 18600, number of used features: 1380
[LightGBM] [Info] Start training from score -5.186526
[LightGBM] [Info] Start training from score -5.320057
[LightGBM] [Info] Start training from score -5.215796
lightgbm python
1 Answers

It looks like you are training a multiclass model. The "prediction" that lightgbm actually keeps track of in binary / multiclass modeling is the log odds (or logit), and not the probability. That's the initial score you're seeing.

user_64415150
Related