Just questions

[IbanezJS]

Hi !

1)Is there 'feature importance' in LightGbm like in XGBoost?
2)What is the fastest way to get prediction for single row on already trained model. If say i have a stream of newcoming data rows and i want to get prediction for each row as fast as possible

[bbennett36]

1: SHAP is my preferred method to find feature importance. You can also use the 'plot_importance' function to visualize the top features by 'gain' and 'split'.

2: Call the predict method and pass it the data that you want to score.

I recommend going through the quickstart.

[IbanezJS]

@bbennett36 Thanks a lot! For SHAP and python links.

2: Call the predict method and pass it the data that you want to score.

Pretty obvious. But i rather looking for special c++ low latency approach to score streaming data. If there are any tricks and specificity.

[guolinke]

@IbanezJS cpp predict function is in https://github.com/Microsoft/LightGBM/blob/master/src/application/predictor.hpp
However, it isn't design for steam data.
If you want speed up the single row prediction by multi-threading, you can change this function:
https://github.com/Microsoft/LightGBM/blob/master/src/boosting/gbdt_prediction.cpp#L9-L27

[IbanezJS]

@guolinke Hi
which part you suggest to go parallel?
for (int k = 0; k < num_tree_per_iteration_; ++k)
this cycle? Are the models_ all independent?

Actually I don't get math here. Why num_iteration_for_pred_ and early_stop need for prediction on already trained model?
How early_stop may decide to stop prediction iterations?
Please give some educational references for this questions.

[guolinke]

you can use openmp.
The models_ are independent.
sorry, not everything has reference.

The pr of early_stopping in prediction : #550

you can specific num_iteration_for_pred_ to use part of trees for prediction, default is using all trees.

BTW, if you are not using multi-class, and didn't use early stopping in prediction, the changes is quite easy:

  double sum = 0.0f;
  #pragma omp parallel for schedule(static) reduction(+:sum)
  for (int i = 0; i < num_iteration_for_pred_; ++i) {
    sum += models_[i]->Predict(features);
  }
  output[0] = sum;

[IbanezJS]

@guolinke thanks for the hint!

does #pragma omp parallel really beneficial?
I mean how heavy model->Predict() could be and what is average of num_iteration_for_pred_?
If compare to threads synchronization and context switch time.

Are there any further simplifications if using only binary classification output?

[guolinke]

@IbanezJS it depends on the complexity of your model, e.g. num_leaves, num_trees. You can just test it by run it.
I think it cannot be further simplified.