Knowledge-Based Construction of Confusion Matrices for Multi-Label Classification Algorithms using Semantic Similarity Measures
This work addresses the need for more semantically-aware evaluation metrics in multi-label classification, offering a domain-specific improvement over traditional statistical methods.
The paper tackles the problem of evaluating multi-label classification algorithms by introducing a method that uses ontology-driven semantic similarity measures to align expected and predicted labels, resulting in more precise confusion matrices for better assessment.
So far, multi-label classification algorithms have been evaluated using statistical methods that do not consider the semantics of the considered classes and that fully depend on abstract computations such as Bayesian Reasoning. Currently, there are several attempts to develop ontology-based methods for a better assessment of supervised classification algorithms. In this research paper, we define a novel approach that aligns expected labels with predicted labels in multi-label classification using ontology-driven feature-based semantic similarity measures and we use it to develop a method for creating precise confusion matrices for a more effective evaluation of multi-label classification algorithms.