Figure 3

Schematic of the process of outlier selection, and the variations in MAEs during outlier selection using KNN classifier. A: Each sequence of Dataset 1 was individually deleted to select the resulting datasets that produce improved learning performances. The top 50 new subsets were kept at each round, and individual deletions were repeated. Due to computational complexity, the best feature set selected by wrapper as described in the paper was used in training. B: Variation of learning performance using KNN classifier during outlier selection. The mean absolute errors of prediction improved with selective removal of putative outlier instances. There was a significant improvement of learning accuracies at Round 2 and 4 (thick lines). After Round 4, the improvement of learning performances slowed down significantly.