On the Normalization of Confusion Matrices: Methods and Geometric Interpretations

The confusion matrix is a standard tool for evaluating classifiers, providing a detailed view of model errors. In heterogeneous settings, its entries are influenced by two main factors: class similarity, reflecting how easily the model confuses certain classes, and distribution bias, stemming from imbalanced training or test distributions. Because confusion matrix values jointly reflect both factors, it is difficult to disentangle their individual effects. To address this issue, we introduce bistochastic normalization via Iterative Proportional Fitting, a generalization of row and column normalization. Unlike standard approaches, this method recovers the underlying structure of class similarity. By disentangling error sources, it enables a more precise diagnosis of model behavior and facilitates classifier improvement. We further establish connections between normalization, importance sampling, and class representations in the model’s latent space, thereby offering a clearer interpretation of normalization schemes. Our implementation is publicly available.