TF - Decision Forests x Multiple Columns as labels

Hello everyone,
I currently working a ensemble with some Neural Networks and love the TFDF too, but the dataset has 8 target columns (labels) which not possible to merge. Despite the binary values of all that columns, merging lose many information because its not possible to predict new values from that dataset.

I not wish use other libraries than TF, if has a solution with YGGdrasil or with more learn, i am totally dispose.

It is a CSV with 35 columns, and 7 labels, but I infer more one, so I want to use NN or TFDF and if it works, do the ensemble, but it’s not mandatory, I just want to compare the results.

Thanks for all community.

Hi @Andre_Galhardo ,

Could you please give us more details/context to help you in better way. What exactly you wants to achieve?


1 Like

Sure, sure.

The Pandas Data-frame is in Kaggle Competition:

He consist in possible 8 labels or target outcomes, and 25 columns, with the objective of the Classification with a metric AUC. However, the labels are in columns separately, not in a one column as traditional multi-class prediction.

Despite the labels are binaries in columns, it’s not possible to convert in only one column, because the predictions are for the total labels.


You can always use DNN for any type of classification problem. This can be better than the standard ensemble methods but requires lots of considerations (depth of the model, no of layers, type of optimizer, activation function) and many more. If you want to compare the results then go for it. DNN gives more flexibility and power to control the whole training process. DNNs are time consuming and require more computational power.

Yes, Ajay_Krishna. I agree to you.

My project for this competition has results above the medium DNN, but I wish can do a baseline with TFDF too and tests some ensembles.

1 Like