Figure 4

Classification performance vs. the size of training dataset. We used T₂₀ to show the impact of training dataset size. At the beginning, only 10% of 2198 sequences in T₂₀ were treated as training samples while others (90%) were used to test miRFam. At each round, we increased the training set by one partition (10%), and accordingly the testing set was reduced by one partition (10%). This process continued iteratively till half of T₂₀ was for training and the other half for testing. The result of normal 5-fold crass validation is also shown.