# Shifted feature distribution across different datasets

I am trying to validate a classifier using two different training and testing datasets.

The feature I am considering is a feature constructed doing the fold-change between two original features, i.e. $log2(feat_A / feat_b)$

The problem is that, across the two different datasets, I see the same distribution of fold-change across groups but with different values. See, in the images attached, the values for each group (each box) in the different datasets (each plot).

What may be causing the differences in fold-change? In this case the fold-change is different also in sign, which means that while in one dataset $feat_a$ is greater than $feat_b$, the opposite is true for the other dataset. However, the pairwise relation between the two groups (the two boxes in each plot) is maintained.

I was thinking about different normalization procedures on each feature across different dataset, which may explain the shift of the distribution of the fold-change. What other problems may be in these datasets?

• Did you stratify your data? Also, sorry but only trying to double check, do you have any sequence numeric I’d field or date feature that might potentially be mislabeled and influence the result? – The Lyrist Jun 12 '18 at 14:25

• In each plot, each boxplot represents a group of samples. My task is to classify each sample correctly in the first or in the second group. If I train the classifier on the first dataset (first plot) I am unable to classify correctly the samples described by the second dataset (second plot). This is because a simple classifier would put a boundary in fold-change $= -2$ to split the distribution of the samples in the first plot (dataset). However, using the same classifier in the second plot (dataset) will fail to classify samples correctly, because all fold-changes are greater than $-2$. – gc5 Sep 20 '16 at 7:26