Skip to content
Better HN
Top
Best
Ask
Show
New
Jobs
Search
⌘K
0 points
psandersen
7y ago
0 comments
Save
Share
DNN's make good features though, especially if you have time series data or lots of text.
I agree that the final model should be a randomforest/xgboost/lightgbm for typical tabular data.
0 comments
1 comments · 1 top-level
top
newest
oldest
mlthoughts2018
7y ago
I meant that extracting an intermediate layer as a feature embedding and then sticking a classical model on top of it
performs worse
than curating features through domain-specific expert tuning, for a ton of diverse application domains.
j
/
k
navigate · click thread line to collapse