Can a heterogenous approach to data normalization yield a better consensus learning result?

Question

[fuellen]

dear All,

The following makes sense intuitively, but I cannot find a reference:

• Let's say I've got 3 datasets, for example differential gene expression, of related experiments, where normalization was done in 3 different ways.
• Then I consider the pathway enrichment results found to be in common for all 3 datasets.

In other contexts it is known that standardizing helps internal validity, but it may diminish external validity. Is that the case here? That is, if I'm lucky to find common pathways while keeping the normalization heterogeneous, the result generalizes better? And if I have to resort to the same normalization process to see any consensus, that consensus is less likely to be true on related yet unseen data??

Is there any work on this, any papers? I could find work on ensemble diversity, heterogenization and causality patterns, but that's all related, not the same thing.

Thanks a lot in advance! G

Answer 1

[LauferVA]

I find this question difficult to answer due to its generality. I suppose I will have to answer in a general way. I don't think it will be helpful, but Ill try to provide something more practical after. TL;DR:

In general, yes. At google, right now, a major focus of theirs is data standardization of many kinds (for their learning projects).

Practical tidbit:

I don't think there is anything inherently wrong with running multiple models or learning approaches for your data, unless you are doing it without a biological rationale for doing so. Without knowing your specific context (and very possibly, even if I did) I think the best/most practical advice I can give you is to try several approaches and try to determine why the performance is better or worse for that specific question and dataset.