If i have 3datasets and make comparison between them and after calculation of pca components for each dataset. For instance, If D1 gave us 7 pc components , D2 gave 5pcs and D3 gave 4 pcs . How can I combine these differences to select specific unique number of pcs for the the three datasets? Which mathematical rule can be used ?
Dear Radwa Elawadi
Follow these steps:
1. Remove an object
2. Scale the remainder.
3. Carry out PCA with a certain number of components.
4. Scale the deleted object according on the parameters found in (2)
5. Use the PCA model to forecast the item.
6. Determine the PRESS for this item.
Kind Regards
Qamar Ul Islam
@Qamar UI Islam
I already applied ckf and calculated the PRESS then measured the reduction in the prediction error till the minimum (threshold) to identify the number of pc components for each dataset. However, how can i combine the different pcs to have only one value of pc components such(4 pcs) for the three terms of datasets.
Will apply reconstruction error??
There is presumably no need to perform cross-validation to select number of PCs. If you, obviously, due to involved mean and covariance, you end-up with different number of PCs each time. However, if you want to evaluate the overall machine learning or data science pipeline, then
- perform k-fold cross-validation on the training dataset or full dataset.
- for each fold after standardization (for example based on explained variance), select the number of PCs.
- Using number of the desired PCs and the corresponding weight matrix, get the score (meta features) and train the classifier.
- Using the mean and variance of training fold, standardize the test fold. Using the weight matrix obtained from the training fold, get the scores (meta features) for the test fold.
- Now test the trained classifier.
- Badges
- Science topic