Does testing a training dataset guarantee successful results?

Question

If I test an image that has been previously used to train a classification model, is it guaranteed to classify correctly?
My guess is that since the parameters have been trained with other images as well, there is no guarantee of getting a correct classification, just a high probability.

Erwan · Accepted Answer

This is correct, there's no guarantee at all, not even a high probability. As usual It depends on the type of model, the data, the number and distribution of the classes.
However there's of course a higher chance that the instance would be correctly classified. That's why one shouldn't use a test set containing training instances to estimate the performance of the model, since there's a high risk the performance would be overestimated (data leakage).

Does testing a training dataset guarantee successful results?

One Answer

Add your own answers!

Ask a Question