There are many points to be looked into I am just adding the few I encountered during my research work:

1] Decide over the output required should be seen accuracy, recall or precision.

2] For which we must follow up with confusionmatrix.

3] Check for MAE, MSE and RMSE

4] Check Outliers and clear it or replace the cases. And understand the reason is it really a human error or some unexpected case

5] Never go for the distribution of training and testing datasets. Go for 3 distribution Training, validation and testing datasets.

6] One must save the earlier model output as training input for a new model.

I am not a good research paper writer but I have explained all the points in my research paper published you can have a look into it

Surface Defect Detection using Convolutional Neural Network Model Architecture

• DOI: 0.55708/js0105014

EDUARD

Outliers identification — esp. in multivariate data — is very longstanding and both informal and formal methods (formal requiring relatively little, or no analyst overview). Re: Classical Formal (available NC online) is: Rocke & Woodruff (1996). Identification of outliers in multivariate data:…” ASA, 91(435). Re Contemporary Formal (available NC online): Yue Zhao (23 May, 2023). “Outlier Detection: Automation, Systems (CMU Dissertation).

ALVAH

Alvah C Bittner, PhD, CPE

EDUARD

I would also strongly suggest looking at the nature of “outliers. “ IME, they may point to

1) Enhanced data collection methods and/or metrics ( respectively, improving future  efforts, but sometimes remarkable improvements in model validities)

2) Breakthroughs in understandings (pointing to new research directions, eg  important genetic polymorphism, and a inanticipated mechanism for reducing disease, transmission, or immediate product improvement opportunity)

ALVAH

 Alvah C. Bittner, PhD, CPE