Goid morning, it is one of issue when i review the literature i did not find article concerning threshold of accuracy. It seems that there are not threshold values for this such as other validation metrics. I consider that a value too close of 50% means that the classification was done randomly and a value too close of 100 % means that the model overfits and there will be issue to run the algorithm with new data.

@Inès how do you identify over fitted model . Because some models can perform well by obtaining values close to 100%

When using the CNN for the purpose of classification, it may reach high levels of accuracy, which may approach 100 percent

Hey Ernest Akpaku ,

I would like to suggest considering all the iterations for classification with respect to the epochs, batch size and the train-test splits. Considering not only the accuracy but also the Precision and F1 scores of the model, we can reduce the chances of selection of an overfitted model as mentioned by Inès François. Also, the classification technique matters for different datasets as it is different for text-based csv data, image data, etc. Overfitting and Skewness can be easily detected by the above-mentioned metrics by implementing a confusion matrix for each model.

Understanding Confusion Matrix, Precision-Recall, and F1-Score | by Pratheesh Shivaprasad | Towards Data Science

https://towardsdatascience.com/understanding-auc-roc-curve-68b2303cc9c5

The above link might help you gain an insight on how to easily detect if a model is skewed.

Thank you.

Goood afternoon, in your model your aim is to conduct a diagnostic of desease. Saying you have a accuracy rate of 97%, however your number of false negative is 37. It means that for 37 patients, you said you don t have the type 2 diabete and in reality there are sick. Inversely, 50 is your false positive, so you take the decision to begin a treatment for people who do not need it. The fiability of classification models for desease diagnostic can be risky and havd to stay only a help tool to the diagnosis requiring the complentary of medical tests. The consequence will be les s severe if your question is, for example, the loyalty of customers.

Inès François, Rahul Pandya and everyone, thanks for the explanations. Really helped

Good morning Ernest Akpaku , there is no specific threshold for a classification method, but the most your accuracy is high the most your model is able to predict well the disease or the phenomenon studied.

One measure to take in consideration is the FN ratio (false negative), where the phenomenon or the disease is not predicted or the subject is normal where it is not the case in real.

An accuracy of 85% and above can be great.

Dear all, i appreciate the great submissions

Good morning

It depends on thec accuracy.Use Conffusion matrix to find the accuracy.