Models are not able to fit the flight delay dataset

[vxbrandon]

Is the accuracy low??

[motiwari]

@vxbrandon yes, about 80%, which is just the same accuracy as predicting the modal class 0

[vxbrandon]

Even lightgbm get that accuracy with max depth greater than 10!

[motiwari]

@vxbrandon I don't think that screenshot makes sense. Is it the same data? If so, that screenshot implies that LightGBM is learning the wrong thing and doing worse than just predicting the modal class, 0, for everything

[motiwari]

I discussed with Jey; this is happening because at each leaf node, 0 is still the majority label, so we always predict 0. The model is learning, it's just underfitting.

To get better accuracy, we need to set max_leaf_nodes=100 and max_depth=1000000