Mathematical Framework for Class Self-correction Through Reverse Validation in Hierarchical Classification

In machine learning research community, it is very common to utilize the results of each level and perform hierarchy based decision making to achieve better classification accuracy in holistic decision engine. As the classifier accuracy for any generic problem statement can’t be assured to be 100%, yet due to multiple factors, automated selection of super-class can give rise to confusions and may lead to wrong decision propagating through subclass classifications. It is also observed that, confidence scores of a classification task plays a major role in selecting or deciding the belongingness of an object to certain class. The current work presents a generic mathematical framework to address the intra-level and inter-level confusions that arises due to automatic selection of classes based on only high confidence scores, interactions of each level in hierarchical classification problem and even redefining a normal classification problem in terms of hierarchical classification, as suitable. The proposed mathematical framework auto-corrects inference errors of the AI models through reverse validation in hierarchical classification (RVHC). The said framework honors the (a) confidence scores of classification, (b) superclass-subclass interaction configuration, and (c) adaptive weight matrix of overlapped subclass components to formulate projection matrix as far as the RVHC is concerned. The proposed work also has proved that the formulated math works well with various kind of critical real-life problem statements in machine learning and deep learning based classification cutting across the technologies of computer vision, speech processing, and nature language processing.