I was wondering if it's normal that having an accuracy of 94-95%, the prediction is always wrong.

I'm new on tensor flow, and i was wondering if it's normal that having an accuracy of 94-95%, the prediction is always wrong.
I've got a binary classification problem, with 2/3 of the rows of one class.
If i train and test the model, i've got 94-95% accuracy, but if i try to predict some samples, the prediction of the class less present is always wrong. On the other hand, the other class is always predict correctly.
The weird thing is that for the samples prediction i use real rows for the correspondent classes.
I'm using the beginner guide's code https://www.tensorflow.org/get ... nners, i change only the dataset.
Is it normal that even having a pretty high accuracy, the model after try 10 or 20 times, is aways wrong only for one class?
Thanks in advance
You already invited:

CHINTAN PATEL

Upvotes from:

if you changed the dataset it could be that you have a "bad" distribution for the samples. If you would have about 90% of your data in one class your baseline would be 90%. So your model having 90% accuracy would still be kind of bad.
If you would test with each class separately the class which is 90% present would be predicted correct but the other class would be labeled as the most occurring one too and be predicted with the wrong label.

If you wanna answer this question please Login or Register