For classification problems, classifier performance is typically defined according to the confusion matrix associated with the classifier. Based on the entries of the matrix, it is possible to compute sensitivity (recall), specificity, and precision. For a single cutoff, these quantities lead to balanced accuracy (sensitivity and specificity) or to the F1-score (recall and precision).
It really depends on your need and dataset characteristics. Classification accuracy makes sense only if your class labels are equally balanced.
Macro, Micro average of performance metrics is the best option along with the weighted average. You can use the ROC area under the curve for the multi-class scenario. You can generalize the actual binary performance metrics such as precision, recall, and f1-score to multi-class performance.
I suggest you follow https://medium.com/apprentice-journal/evaluating-multi-class-classifiers-12b2946e755b
https://www.datascienceblog.net/post/machine-learning/performance-measures-multi-class-problems/
Please visit the attached links
https://medium.com/apprentice-journal/evaluating-multi-class-classifiers-12b2946e755b
https://www.datascienceblog.net/post/machine-learning/performance-measures-multi-class-problems/
In the real world, classification accuracy based on probabilities is probably much less important than the consequences attached to both correct and incorrect classification. Consider this scenario: you're on an African veldt, and you see a tan shape at some distance. Is it a rock? Is it a lion? If your conclusion, based on probabilities, is that the shape is a rock, you might die.
Can we figure out a metric which takes into account the relative importance for each class, of both correct and incorrect classification?
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification