In my opinion all elements what you said for evaluation the performance of DL models, its very good to judge robustness of your model, all this what are you said is called rapport of classification. we can in some time plot the Confusion Matrix for visualized the confusion between the classes.

I hope that be Clair for you.@ Ibrahim mohamed Gad

you should better balance your data first.

.

you could consider the balanced error rate

Article Classification Assessment Methods: a detailed tutorial

there is no such thing as an absolute measure of performance ... at the end of day, it all depends on your application ; for instance, you could know the costs for TP, FP, TN and NN : in such case, the expected cost should be your prefered metric ...

.

Please use confusion metrics. That one is recommended and easily accepted

Consider using specificity and sensitivity in addition to other metrics

for the imbalance data, the confusion metric is the best metric in which false positive and false negative errors easily observed.

Use Kappa statistic, works best with imbalanced datasets.

Consider performing a pre-processing step to ''balance'' the dataset such that bias in the network performance is eliminated or at least reduced. This can, for example, be done by removing some of the datasets corresponding to examples with higher proportion. I would think it is better to train the network with a small-size balanced data than a large unbalanced data.

If you are dealing with multi-class dataset, you may consider different versions of the metrics such as macro, micro, weighted etc.

For imbalanced class problems:

https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html

There are many techniques to balance your data. After that, for evaluation, you can try k-fold cross validation.

For an imbalanced data, use precision and recall.

For a balanced data you can either precision/recall or accuracy.

To avoid misleading results, it is a good practice to balance an imbalanced data first.

For imbalanced data, you can use weighted classification, then using the weighted average recall (WAR).

Or using unweighted average recall (UAR) for imbalanced data classification.

you can't rely on accuracy measure when you have imbalanced data because it might be deceiving .

but you should get a high accuracy anyway to prove that your model is working well .

then comes the most important measures : precision , recall and f1score

your model is making a lot of false positive predictions that's why your precision is not hight (that's not a good thing )

your model is not making a lot of false negative predictions that's why your recall is higher (that's the good thing )

f1 score is the harmonic average ( keep in mind it's not a normal average it gives weight to either precision or recall depending on something called beta value )

your f1 score is good because it's automatically gives a little weight to your recall (which is already good ) .

for me i think your precision is not good and that can't be overlooked .

please , correct me if i'm wrong .

In my opinion all elements what you said for evaluation the performance of DL models, its very good to judge robustness of your model, all this what are you said is called rapport of classification. we can in some time plot the Confusion Matrix for visualized the confusion between the classes.

I hope that be Clair for you.@ Ibrahim mohamed Gad

I would recommend using Kappa score.

My recently published paper with the title of “Applying Separately Cost-sensitive Learning and Fisher's Discriminant Analysis to Address the Class Imbalance Problem: A Case Study Involving a Virtual Gas Pipeline SCADA System” and URL of https://www.sciencedirect.com/science/article/pii/S1874548220300214 may be helpful.