Sample/label weighting to handle class imbalance

[severinsimmler]

Resources:

• https://docs.allennlp.org/models/main/models/tagging/models/crf_tagger/
• Handling unbalanced datasets in the CRF tagger allenai/allennlp#4619 (comment)
• https://eraldoluis.github.io/2022/05/10/weighted-crf.html

[severinsimmler]

Maybe via loss function?

See also: https://www.diva-portal.org/smash/get/diva2:1680166/FULLTEXT01.pdf

And also:

chaine/chaine/_core/crfsuite/lib/crf/src/train_l2sgd.c

Lines 179 to 197 in 3a2626b

	/* Loop for instances. */
	sum_loss = 0.;
	for (i = 0; i < N; ++i)
	{
	const crfsuite_instance_t *inst = dataset_get(trainset, i);
	/* Update various factors. */
	eta = 1 / (lambda * (t0 + t));
	decay *= (1.0 - eta * lambda);
	gain = eta / decay;
	/* Compute the loss and gradients for the instance. */
	gm->set_weights(gm, w, decay);
	gm->set_instance(gm, inst);
	gm->objective_and_gradients(gm, &loss, w, gain, inst->weight);
	sum_loss += loss;
	++t;
	}