PCA is applied separately to training and test sets to compute explained variance for each. A data scientist notices the explained variance ratios differ significantly between splits. What does this indicate?