I am training an AutoML data model on a set of power plant output data. In this case, AT is the ambient temperature (and one of the features), and the target is PE (the power output by the plant).
I am in the process of creating a VertexAI pipeline run, and in the AutoML "Training options" section, I have requested the calculation of the correlation between each feature and the target.
The correlation given between AT and PE is 1.157. However, this is supposed to be a Cramer's V number, and therefore somewhere between 0 (no correlation) and 1 (perfect correlation). The tooltip for the correlation column says "The Cramér's V correlation statistic between this column and the target column. This ranges from zero to one, where zero indicates no correlation and one indicates perfect correlation. A low correlation suggests that the column can be excluded from the model without much performance penalty. An unusually high correlation is indicative of target leakage and/or a categorical feature with a very high cardinality relative to number of rows."
So how can the correlation value given for the AT feature be above 1? Has anyone seen this before?
I have done a very basic linear regression between AT and PE in Google Sheets and it's giving me an R2 of 89.89%, so the correlation is indeed high, but I don't know why the Cramer's V given by Google Cloud is above 1. There are no missing values for AT or PE in my data.