This study reference evaluates diagnostic frameworks designed for high-dimensional multi-class and multi-label classification environments. It maps out how standard binary confusion matrices scale into isolated One-vs-Rest (OvR) sub-matrices and comprehensive evaluation summary sheets.
1. Multi-Class vs. Multi-Label Model Spaces
Multinomial Logistic Regression: An extension of classical binary logistic models used to predict the conditional probability distribution across three or more mutually exclusive target categories. Each row instance is restricted to a single class assignment.
Multi-Class Confusion Matrix: A consolidated C x C contingency table (where C equals total target classes) cross-tabulating factual target categories against model projections, mapping directional misclassifications across classes.
Multilabel Classification Space: An advanced environment where individual rows can belong to multiple distinct classes simultaneously, eliminating mutually exclusive constraints. Instead of a single model output, predictions track as an array of binary flags.
2. The Localized One-vs-Rest (OvR) Matrix Engine
Evaluating multi-label frameworks requires decomposing the problem space. A Multilabel Confusion Matrix constructs an array of isolated 2x2 binary contingency matrices—one for each unique class—using a One-vs-Rest approach. For any single targeted class, test statistics are defined locally:
Class-Specific Cell Quantifiers
True Positive (TP): The count of instances correctly predicted by the model as belonging to the active target class.
True Negative (TN): The count of instances correctly predicted by the model as not belonging to the active target class.
False Positive (FP): The count of instances incorrectly predicted by the model as belonging to the active target class when the actual ground truth is negative.
False Negative (FN): The count of instances incorrectly predicted by the model as not belonging to the active target class when the actual ground truth is positive.
3. The Comprehensive Classification Report Index
A Classification Report is a structured diagnostic readout summarizing performance across high-dimensional target spaces, documenting local class statistics alongside global macro and weighted averages:
Local Class Metrics
Precision: The proportion of true positive predictions relative to all instances flagged by the model for that specific class:
Precision = TP / (TP + FP)
Recall (Sensitivity): The proportion of true positive predictions relative to the absolute universe of actual ground-truth members for that specific class:
Recall = TP / (TP + FN)
F1-Score: The harmonic mean balancing precision and recall for that specific class, providing a robust fit measure independent of sample size:
F1-Score = 2 * ((Precision * Recall) / (Precision + Recall))
Support: The baseline count of factual ground-truth instances belonging to that specific class within the active test dataset, serving as a sample volume contextual anchor.