Figure 4

Performance of SVM and naïve-Bayes classifiers. The performance of the SVM for identifying interaction abstracts was evaluated using 10-fold cross-validation on a set of 1094 abstracts. The performance on this task is measured in precision and recall. There is an implicit tradeoff between precision and recall that can be varied if the decision boundary is set to some value other than 0. In this evaluation, when the decision boundary for the SVM is set to 1, recall and precision are 0.57 and 0.99 respectively. When the decision boundary is set to -0.99, recall and precision are 0.997 and 0.71 respectively. Finally, if the decision boundary is set to zero then precision and recall are both 92%. In other words, when the decision boundary is set to zero and the SVM is applied to all abstracts in PubMed, it will miss approximately 8% of interaction documents (recall) and 8% of the identified interaction documents will not be interaction documents (precision). Under similar conditions, the naïve-Bayes classifier described here would only have a precision and recall of 87%.