Automl on data with nulls

[matt7salomon]

I am using Automl to only train Deeplearning algorithms on a dataset which is sparse and has a lot of nulls. I preferably shouldnt impute this data because nulls really have a meaning as the customer had not interacted with that feature. Automl results are terrible on this dataset. It doesnt even predict in ballpark of the output. Any suggestions?

[tomasfryda]

I would suggest trying other algorithms, namely the tree-based ones since they can handle missing values as separate category (e.g. GBM, https://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/gbm-faq/missing_values.html). You can also keep AutoML unrestricted to find out which model type suits those data the best.

[matt7salomon]

We already have a non h2o lightgbm model in place which works just fine and not perfect. We wanted to try deep learning.