Abstract
In addition to pursuing accurate analytics, it is invaluable to clarify how and why inaccuracy exists. We propose a transparent classification (TC) method. In training, data consist of positive and negative observations. To obtain positive patterns, we find the intersection between each of the two positive observations. The negative patterns are obtained in the same manner. Next, pure positive and pure negative patterns are established by selecting patterns that appear in only one type. In testing, such pure positive and pure negative patterns are used for scoring observations. Next, an observation is classified as positive if its positive score is not zero or if both its positive and negative scores are zero; otherwise, it is classified as negative. By experiment, TC can identify all positive (e.g., malignant) observations at low ratios of training to testing data, e.g., 1:9 using the Breast Cancer Wisconsin (Original) and 3:7 using the Contraceptive Method Choice. Without fine-tuned parameters and random selection, the uncertainty of the methodology is eliminated when using TC. TC can visualize causes, and therefore, prediction errors in a network are traceable and can be corrected. Furthermore, TC shows potential in identifying whether the ground truth is incorrect (e.g., identifying diagnostic errors).
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have