Table 2. Benchmark of the abilities of classifiers to distinguish different ions from one another in the blind, curated test set.

Shown here are the unmodified decision-tree algorithm as well as SVMs trained on the curated training set using all features of a site and the same SVM with a simple filter applied to the predictions. The pairs of numbers for each SVM and ion pair indicate the number of correct identifications of the first and second ion, respectively.

		Total sites		Decision tree		All features		All features (filtered predictions)
Ion 1	Ion 2	Ion 1	Ion 2	Ion 1	Ion 2	Ion 1	Ion 2	Ion 1	Ion 2
Ca2+	Mn2+	108	70	60, 0	31, 2	86, 9	59, 8	84, 5	61, 10
Ca2+	Fe2+/3+	108	17	60, 0	9, 0	92, 0	17, 0	90, 0	14, 1
Ca2+	Ni2+	108	49	60, 0	32, 0	91, 0	48, 1	90, 0	38, 2
Ca2+	Zn2+	108	64	60, 0	47, 0	92, 0	60, 0	90, 0	52, 1
Mn2+	Fe2+/3+	70	17	11, 1	6, 0	51, 2	15, 17	50, 2	12, 16
Mn2+	Ni2+	70	49	22, 0	28, 0	52, 2	46, 16	51, 7	36, 15
Mn2+	Zn2+	70	64	26, 0	33, 0	46, 3	57, 22	45, 5	50, 21
Fe2+/3+	Ni2+	17	49	2, 0	27, 0	11, 5	43, 6	8, 9	33, 6
Fe2+/3+	Zn2+	17	64	2, 0	35, 1	11, 2	58, 6	9, 3	50, 5
Ni2+	Zn2+	49	64	1, 0	24, 0	18, 5	55, 30	17, 3	49, 21
Total true positives				569		1008		929
Total false positives				4		134		132