Large-Scale Coarse-to-Fine Object Retrieval Ontology and Deep Local Multitask Learning

Algorithm 5

Finding best thresholds for multilabels.

	Procedure: FIND_BEST_THRESHOLDS
	Input:
	trainPred //an array of multilabel predictions of each sample in the train set
	trainLabels //an array of multilabels of all samples in the train set
	initThresholds //an array of candidate thresholds with value in [0, 1]
	Output:
	bestThresholds //an array of best thresholds of each label
	Goal:
	Find best thresholds for multilabels in an MCC imbalanced data problem solver
	BEGIN
	for i = 0 to numberOfLabel do:
	begin
	[possibleMattVals] ⟵ [NULL]
	#choose threshold
	for j = 0 to length (initThresholds) do:
	begin
	[pred] ⟵ [NULL]
	[currTrainLabel] ⟵ [NULL]
	#get prediction from current threshold
	for k = 0 to numberOfSample do:
	begin
	if trainPred [k][i] >= j:
	pred ⟵ 1
	else:
	pred ⟵ 0
	currTrainLabel ⟵ trainLabel [k][i]
	end
	possibleMattVals ⟵ matthrews_corrcoef (currTrainLabel, pred)   #visit Equation (7) for full computation of MCC.
	end
	bestThresholds ⟵ initThresholds [position of max(possibleMattVals)]   #the threshold with highest MCC value in the initialized threshold list for each attribute is chosen.
	end
	END