Max-Bladen
commented
2 years ago Hi @aljabadi,
I need a bit of advice for this one. I've successfully implemented a way to retain user specified features in sparse functions via the new retain.feats parameter. It can take either indices or names of the desired features to retain and is implemented for spca, spls, splsda, mint.spls, mint.splsda, block.spls and block.splsda.
I'll refer to features selected by the sparse method (via keepX) as "selected" features. Those specified by the user (via retain.feats) will be referred to as "retained".
Here's the problem: the loading values of these retained features is severely overestimated. This is due to the selected and retained having their loadings reduced by max(abs.a[!select_feature]) (look here). Using the below code as an example:
data(liver.toxicity)
X <- liver.toxicity$gene[, 1:200]
Y <- liver.toxicity$clinic
colnames(X)[6:10] <- c("A", "B", "C", "D", "E")
colnames(Y)[8:10] <- c("A", "B", "C")
retain.feats <- list(X=6:10,
Y=8:10)
spls.obj <- spls(X, Y, keepX = c(6,1), keepY = c(6,1), retain.feats = retain.feats)So, the features to retain are named A -> E so they can be easily identified. In a given iteration, the loading values of the selected and retained features are as follows:
A B C D E
37.56952 13.64893 20.92075 3.42440 20.26891
A_42_P649672 A_43_P11568 A_43_P21626 A_42_P454114 A_42_P683537 A_43_P21372
63.32374 68.51839 66.69514 66.56536 67.45634 66.31351Unsurprisingly, the selected features (starting with A_4*) have the maximal absolute loadings, while the retained features have a range of values. These are then each scaled down by the maximal loading of all non-selected features (max(abs.a[!select_feature]) = 62.85361). Now the values look like:
A B C D E
-25.2840898 -49.2046839 -41.9328549 -59.4292085 -42.5846995
A_42_P649672 A_43_P11568 A_43_P21626 A_42_P454114 A_42_P683537 A_43_P21372
0.4701328 5.6647843 3.8415326 3.7117548 4.6027321 3.4599011Now if we look at the final loading values you can see the overestimation of the retained features (particularly on the X dataframe).
plotLoadings(spls.obj)
Created on 2022-06-22 by the reprex package (v2.0.1)
Mathematically, these values are correct but the scaling results in their inflation. I played around with removing the - max(abs.a[!select_feature]) but this obviously resulted in different values for other components of the method (eg. weights and AVE).
My primary question is, seeing as the loading values are scaled down via the L2 norm (sqrt(crossprod(loadings)) - look here) is the - max(abs.a[!select_feature]) necessary? If not, then the removing this subtraction resolves the loading inflation issue, like so:
Created on 2022-06-22 by the reprex package (v2.0.1)
Seeing as the loading values (as well as other components, eg: AVE) change to differing degrees when we remove this subtraction, I'll assume that this subtraction is necessary. Hence, where should I go from here? Is there a better way to implement the retain.feats functionality without this issue? Should I just forget about this feature?
Cheers
mixOmicsTeam
Users requested some time ago to include a feature within our sparse methods which allows for them to pass in a specific set of variables to be guaranteed to be included in the model. Relevant user facing functions (
spca,(mint).(block).spls(da)) all now take theretain.featsparameter to allow this funcitonality.Check.retain.feats()handles checks and pre-processing of the inputretain.feats. This parameter influences the regularisation that occurs withinsoft_threshold_L1()called bysparsity(). Along withselect_feature, all features included as part ofretain.featsdo NOT have their loadings reduced to 0.