As tpot seems to rely solely on scikit-learn for (meta-) estimators the lack of extended multi-label classification strategies is quite noticeable. The work on some of the strategies and algorithms is stalled for quite some time now in scikit-learn (https://github.com/scikit-learn/scikit-learn/pull/2461 (label powerset) and https://github.com/scikit-learn/scikit-learn/pull/3727 (classifier chains)). As such there is work being done on scikit-multilearn and it already brings at least some novel working algorithms and strategies.
What do you think of including scikit-multilearn (at least for the time being) to extend the support of multi-label classification?
(For clarification: multi-label classification is defined as finding a subset of predicted labels out of a total label set, i.e. Y_hat = {1,3,5}, meaning multiple "classes" (or labels in this context) are assigned to one sample.)
Hey there,
As tpot seems to rely solely on scikit-learn for (meta-) estimators the lack of extended multi-label classification strategies is quite noticeable. The work on some of the strategies and algorithms is stalled for quite some time now in scikit-learn (https://github.com/scikit-learn/scikit-learn/pull/2461 (label powerset) and https://github.com/scikit-learn/scikit-learn/pull/3727 (classifier chains)). As such there is work being done on scikit-multilearn and it already brings at least some novel working algorithms and strategies.
What do you think of including scikit-multilearn (at least for the time being) to extend the support of multi-label classification?
(For clarification: multi-label classification is defined as finding a subset of predicted labels out of a total label set, i.e.
Y_hat = {1,3,5}
, meaning multiple "classes" (or labels in this context) are assigned to one sample.)