I have a question about the underfitting and the overfitting in ML.
Now, when the accuracy of the training is more than the accuracy of testing, this is called the overfitting, in this case, we can not trust this model in the classification or clustering , but when the accuracy of the testing is more than the accuracy of training, what can be called this case scientifically and is it considered positive in the work of the model or not?
Addition to that, how we can to detect the underfitting from knowing only the accuracy of training and testing?
Thanks in advance.
Underfitting means that you can't even fit a training set = training set accuracy is low. Is your testing set big enough?
Hi Manar
Try to do the following things :
1- May by you have too many features, i think you need to increase your data or decrease your number of features.
2- try to change splitting of your data for training and testing, split the dataset to be 90% training and 10% test data.
3- Use cross validation
Also, please there is a researcher asked the same question you asked, kindly follow this link:
https://www.researchgate.net/post/How_is_it_possible_to_obtain_better_results_on_the_test_set_than_on_the_training_set
Thank you all very much.
Dr. Mohammed,thanks for the discussion link.
I have solved the problem that related to the high accuracy in the training set (The overfitting), and now the accuracy of the test set is higher than the training set. (What is called this case?)
And, how do I know that the model has reached to fitting in classification?
I don't think that there is a specific name for the case that the test set accuracy is higher than the training set accuracy. Usually it is the case when the test set is much easier than the training set. As Mohammed mentioned, you might want to perform a cross-validation by obtaining different training and test sets by partitioning the whole dataset several times. You will see that the test set accuracy will go down. If not, please check your evaluation.
I'd recommend to look at Fig 5.3 here https://www.deeplearningbook.org/contents/ml.html to appreciate the difference between Underfitting and Overfitting. Also, as in the same Fig, I'd expect to naturally observe a higher training accuracy than test accuracy, which isn't bad per se.
How to calculate tp, fp, tn, fn for trained data, in order to get accuracy, in R ....
Shahid Parwez You should open a new question instead of replying to a barely related one-year-old discussion with a question. You'll have better chances to get an answer.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification