Fig. 3: The Output of PDD for Cytochrome C APC data.
From: Theory and rationale of interpretable all-in-one pattern discovery and disentanglement system
a It is part of the input data showing the entanglement of patterns associated with different classes/sources within entities. A synthetic pattern S73 = T, S6 = Y, and S88 = F are implanted on a Mammal, a Fungus, and an Insect. b It shows PDD separating the statistically correlated patterns occurring on entities in the original data into disentangled patterns pertaining to distinct sources/classes. c It shows the summarized Knowledgebase with class label given. Note that the union of the disentangled patterns discovered in each DSU occurred only on entities related to one class/source—indicating perfect disentanglement. d In the Knowledgebase from no-class-label dataset, the knowledge space shows that the union pattern in DSU[1 1 1] is shared by Mammal and Fungus, revealing their common functionality. However, when PDD was applied to these entities, patterns associated with Mammal are separated ((c) in Supplementary). In (c), PDD discovered 5 outliers. In (d), unaffected by class, we found 3 entities formed a rare group with the implanted pattern [T Y F] in DSU[3 1 1]. EID84, found also as Fungus (as labeled), we assigned that as its final status by Rule 1. PDD attained a class association accuracy of (95-3)/95 = 97.83% after result integration. e In the abridged entity cluster, each row is an entity. The PDD Results columns show the DSU of the discovered pattern(s), followed by its class label and the final classification results (c), and the number and the class status of all the entities placed in each cluster after the readjustment. The last column displays the cluster placement status - correct (cor), readjust (after Cra), and Inc when the entity is placed into a wrong cluster. After class correcting (Cra), only EID70 and EID73 were found misplaced. Since EID73 was originally labeled as Fungus but found to contain patterns associated with Plant instead, it was considered as a Plant. Hence, with only one error, PDD obtained a placement accuracy of 98.91%.
