search

byu-dml / metalearn

BYU's python library of useable tools for metalearning
MIT License
22 stars 6 forks source link

Feature/instance-level metafeatures #93

Open bjschoenfeld opened 6 years ago

bjschoenfeld commented 6 years ago

All of our metafeatures are computed in aggregate accross all a given dataset. D3M allows for metafeatures that are more granular, i.e. feature/column based, instance/row based, or even value/cell based. We already compute some row- or col-level metafeatures, but discard that information in aggregation. In other words, this is easy to return.

However, is this helpful? Are there uses for these metafeatures? It is not clear to me how they could be used in metalearning applications, since these are dataset-dependent metafeatures and not comparable between datasets.

joaquinvanschoren commented 6 years ago

Sure. Maybe you want to build a histogram instead of the min/max/mean, or detect whether there are outlier values. At least for col-level meta-features. More generally, users may want to experiment with all kinds of new ideas about how to leverage them.

emrysshevek commented 5 years ago

So I added some of the feature level metafeatures rather hastily and the output format is inconsistent with the other metafeatures we have. I think it might be best for now to remove them and add them back in when we have better methods for grouping or have a better idea of how to return them.

epeters3 commented 5 years ago

@macetheace96, like we talked about offline, what we're doing for the time being is rolling back these changes for the upcoming D3M submission, then readdressing this issue in the future. Thanks for being willing to do that.