Why is the R-value 0.95 for raw data but drops to 0.85 for normalized data in ANN?

Why is the R-value 0.95 for raw data but drops to 0.85 for normalized data in ANN ? does it happen or am I doing something wrong

Answer 1

[akshatsood]

Hi @Sunita,

I understand that you are observing an R-value of 0.95 for raw data, which drops to 0.85 for normalized data in your Artificial Neural Network (ANN). I would like to clarify whether these values are from the training dataset or the testing dataset or the validation dataset.

In general, the higher R-value associated with raw data might indeed suggest overfitting, potentially stemming from inadequate feature scaling. By normalizing the data, you are likely achieving more effective feature scaling, which can help prevent overfitting. Consequently, the decrease in the R-value for your Artificial Neural Network model when using normalized data could signify a more robust and generalizable model. This shift in R-value may indicate that the model is now better at capturing underlying patterns within the data, leading to improved generalization.

I hope this helps.

[Sunita]

Here is the R-value for both

[akshatsood]

Hi @Sunita,

Thank you for providing the plots. Upon reviewing them, I noticed an accumulation of data points at the origin in the raw data. This accumulation appears to introduce a bias in the model. Upon normalization, these points seem to significantly alter the data distribution, as evident from the shared plots. It is apparent that the distribution of the data undergoes a noticeable shift due to these points.

Another noteworthy observation is that in the case of raw data, the points appear to align closely with the fit. However, upon normalization, where the scale is adjusted and the range is narrowed down to [-1, 1], these points become notably distant from the 'fit'. This discrepancy may be contributing to the low R-value you are observing.

I would suggest treating the aforementioned points as outliers and then proceeding to train your ANN. By doing so, you may observe improvements in the results. Addressing these outliers could potentially lead to a more accurate and reliable model.

I hope this helps.

[Sunita]

Thanks for ur insight. any idea how can I locate the outliners on the dataset

[akshatsood]

You can eliminate the data points located at or near the origin by using logical indexing. This method involves extracting a specific portion of the data based on defined conditions, effectively filtering out the unwanted points. Additionally, I recommend reviewing the following articles that go deep into identifying and cleaning outliers.

https://in.mathworks.com/help/matlab/data_analysis/data-smoothing-and-outlier-detection.html

https://in.mathworks.com/help/curvefit/removing-outliers.html

https://in.mathworks.com/help/matlab/ref/cleanoutlierdata.html

I hope this helps.