Lightgbm: Init score implementation

[saleepeppe]

I am using lightgbm for modelling a variable which is poisson distributed.
For this variable I have a bias provided by the logarithm of "e", where "e" is one of the predictors.

If I have understood properly, I could set the init score as the logarithm of "e" to train lightgbm.
Unluckily I cannot use any init score for the predictions since the predict() method does not support the Dataset object.

I wonder how is this "init score" is implemented? And why there is no need of using it for the predictions?

Thanks in advance

[chivee]

basically, lightGBM will train from the residual of a certain dataset in each round of training to build a weak tree model, and the init_score indicates the first round of these residual values.

practically speaking, the `init_score' was used in the scenario that you need further training your model against some dataset, and use it to represent the previous model.

[guolinke]

For predictions, you should add initial scores by yourself

[saleepeppe]

For predictions, you should add initial scores by yourself

If I am understanding properly, the scores I obtain from the predictions of a model trained with init_score would be:
init_score + predictions = final_predictions
Is it correct?

[guolinke]

yeah, and you should use the raw_score in prediction. and do your own sigmoid/softmax transform if in binary/multi classifications?

[JYLFamily]

yeah, and you should use the raw_score in prediction. and do your own sigmoid/softmax transform if in binary/multi classifications?

不好意思，关于这个问题想再请教一下。
1、init_score 使用之前学习器的预测的类别还是类别对应概率？
2、final_predictions = sigmoid(lgb 的 raw_score + 之前学习器的 raw_score ) ？
谢谢。

[guolinke]

@JYLFamily

1. init_score is the raw score, before any transformation.
2. yes

[JYLFamily]

@JYLFamily

1. init_score is the raw score, before any transformation.
2. yes

再请问一下，lgb.predict(raw_score=True) == lgb 的 raw_score + 之前学习器的 raw_score 还是 lgb.predict(raw_score=True) == lgb 的 raw_score

十分感谢。

[guolinke]

@JYLFamily The prediction result doesn't include the init_score, you should add it by yourself.