Parameter importance for categorical variables


I have a conditional variable (True/False) in my sweep for freezing a layer or not. It gets quite high importance and a high positive correlation. How do I know if the “True” or the “False” is the “higher” value? In other words, how do I know which to set it to? Is it based on the order specified in the sweep-config? Unfortunately I cannot see any clear pattern in the runs due to a lot of other parameters also being varied at the time in my Bayesian sweep.

Thank you!

Hi @styrbjornkall , I did a brief review of the sweep parallel coordinates chart in your freezing_sweep project, and based on the results, both True and False appear to have equitable influence on your results. It is difficult to interpret boolean results, yes, and our parameter importance panel won’t tell you which value is the “higher” value, just that relative importance of the hyperparameter in respect to the chosen metric.

HI @styrbjornkall , since we have not heard back from you we are going to close this request. If you would like to re-open the conversation, please let us know!