A Preliminary Study on the Prediction of Human Protein Functions - Foundations on Natural and Artificial Computation - page 340

Fig. 4. Histogram of matched protein ranks for the tournament learning strategy

Fig. 5. Histogram of matched protein ranks for the multi-label learning strategy

input layer, 70 in the hidden layer and 359 neurons in the output layer. Learn-
ing parameters were set with the default values defined in the software package.
During cross-validation trials we used a validation set to avoid the over-training
phenomenon. The results were:

- Average precision = 63.4% ( ± 0 . 5%)

±

0 . 9)

- Average rank = 49.0 (

Concerning the independent testing set, Figure 5 illustrates the histogram of
matches with respect to their rank. For each protein with the first 50 ranked
terms we had 79 matches, corresponding to average recall equal to 37.8% and
average precision equal to 3.6%. Therefore, the prediction performance was worse
than that obtained with the tournament learning strategy.