Understanding LabelModel and cardinality

[paulachocron]

I'm seeing some strange predictions on my data. I have labels 0 and 1. I defined 5 labeling functions which sometimes abstain.

When I train a LabelModel with cardinality=2 (the real value), it predicts always 0. I see points where 4 lfs return 1, one abstains, and the predictions is 0. Even points where all lfs return 1, yet the model predicts 0. When I set cardinality=3 results seem much more reasonable. But it's not the real value!

Any thoughts on what could be happening?

This is my LFAnalysis(L_train).lf_summary()

	Polarity	Coverage	Overlaps	Conflicts
0	[0, 1]	0.708946	0.708902	0.501723
1	[0, 1]	0.884765	0.884285	0.615824
2	[0, 1]	0.872596	0.871811	0.601038
3	[0, 1]	0.830985	0.830811	0.602303
4	[0, 1]	0.450735	0.450735	0.355084

[paroma]

You can try changing parameters like class_balance and regularization to see whether that gives better results (see this tutorial and documentation for examples).

We can also help debug further here if we have access to the L matrix for this data.

[github-actions[bot]]

This issue is stale because it has been open 90 days with no activity. Remove stale label or comment or this will be closed in 7 days.