Custom loss function for the multiple responses/outputs

[junjiwoo]

To be more simple, bivariate case, y_1 is 0 or 1, y_2 is continuous. Is it possible to customize the loss function to L = Gini + SSE? Or L = classification lossVregression loss where V can be a matrix for higher dimension.

[guolinke]

you can base on multi-class template, and write the customed objective functions.

[junjiwoo]

@guolinke Thank you for your reply. I am writing to seek further assistant from you about this issue. I found that the lgb.Dataset store label (y) as univariate. Multiclass is still univariate ( unique(y) = 0, 1, 2 ) case. In my case, it is multivariate/multi-output ( for bivariate case y = (0, 100), (1, 200), (0, -10), ...etc ). Thank you for your help.

[guolinke]

Hi @junjiwoo
I think you can use python package to achieve this, by the following processing.

1. construct two lgb.Dataset, the X is the same, but with different label.
2. construct two lgb.Booster, each accepts one lgb.Dataset
3. write two objective functions, each for a lgb.Booster
4. use lgb.Booster.update https://lightgbm.readthedocs.io/en/latest/Python-API.html#lightgbm.Booster.update , to train the booster. You can call this in the loop for mulitply iterations.

BTW, for the step 4, you may need to some shares between two datasets, as your objective function needs the information from both two data.

[junjiwoo]

Thank you for your prompt response. I will follow your advice. I saw many related questions and requests. To the best of my knowledge, nobody implemented this yet. (gbm, mvtboost, xgboost, h2o)

[StrikerRUS]

We already have an issue related to multi-output: #524. I think it's better to keep the discussion in one place.