Distribution of Watermark not linearly changing?

lillijo / xai-evaluation-master-thesis

code for experiments of Lilli Joppien's master thesis: Are We Explaining the Data or the Model? Concept-Based Methods and Their Fidelity in Presence of Spurious Features Under a Causal Lense.

GNU Affero General Public License v3.0

0 stars 0 forks source link

Distribution of Watermark not linearly changing? #12

Closed lillijo closed 6 months ago

lillijo commented 9 months ago

when using normal distribution [0.5, 0.1] for noise terms the curve of rectangles with vs ellipses with watermark looks like this:

but when using uniform distribution [0,1] it looks not linear...

Is this a problem? is the "biasedness" still linear?

lillijo commented 9 months ago

What we are interested in is Pearson Correlation between watermark and shape (that is: the phi-coefficient for binary variables

lillijo commented 9 months ago

result is: