Fig. 3

The result of dropout experiments. a-d show the Cohen’s Kappa score changing with dropout rate under different hash length (32, 64, 128, 256). The x-axis of each figure is the dropout rate and the y-axis of each figure is Cohen’s Kappa score. It is reasonable that the performances of three algorithms all decrease as the dropout rate increase. We can see that DenseFly always outperforms others and has a stable ‘platform’ range where Cohen’s Kappa score decreases slowly when the dropout rate is small, particularly when hash length = 256. The experiments show DenseFly is robust when the dropout event occurs. It should be explained that the original data without dropout has 45% zero elements in the expression matrix, meaning that SIM III-5 dataset (dropout rate = 53.6%) is extremely sparse (over 98% elements is zeros), so all algorithms perform poorly because little information remained in the dataset