[long] the problem is something like this
when you approximate a continuous distribution by sampling
all the samples have the same score
(if you do the aggregation where score = frequency of sampled value)
if you then visualize this, points will tend to overplot and it will be hard to see the real signal without intelligently setting opacity levels
we can get around this for base continuous distributions by just using the value from the primitive scorer
instead of the frequency score
[ndg] ah!
makes sense
[long] but for continuous marginals, we’re hosed
[ndg] yeah.... at some point (maybe soon?) we'll add kernel-based aggregator, which will help with this.
[long] ah, good to know
[robert_hawkins] that’ll be great
[long] at the moment, i haven’t finished fixing the problem for even base continuous distributions yet. hence the advice to distrust monochromatic scatter plots
Reproducing a Slack conversation: