You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Currently, for k routes and top_k most relevant utterances, the final route classification is determined by class-wise similarity aggregation (SUM) and then selecting which class has the highest aggregated similarity value.
Problem
In many tasks, doing this type of SUM aggregation might be bad, as it is very dependent on how the routes are defined which aggregation method makes the most sense.
For very high values of top_k, I would expect the most optimal class to be the one that has the highest consensus across all utterances class-wise. Statistically, I would expect this to be the MEAN similarity value.
Then again, using MEAN could be bad as well, if the utterances for one routes are extremely similar and then for another one they are more heterogeneous. So for the current HybridRouteLayer, it seems evident that which method that is used is dependent on how the end user has defined the routes.
Proposed solution
Allow developers to set aggregation themselves to allow them to find which solution works best.
This solution would be of relevant both for the RouteLayer and HybridRouteLayer classes.
EDIT: I can draft a PR to add this feature.
The text was updated successfully, but these errors were encountered:
Note that I have added this support to both RouteLayer and HybridRouteLayer.
Below is an example on how one will be able to set the aggregation method (supported types are "sum", "mean", and "max").:
Current solution
Currently, for
k
routes andtop_k
most relevant utterances, the final route classification is determined by class-wise similarity aggregation (SUM) and then selecting which class has the highest aggregated similarity value.Problem
In many tasks, doing this type of SUM aggregation might be bad, as it is very dependent on how the routes are defined which aggregation method makes the most sense.
For very high values of
top_k
, I would expect the most optimal class to be the one that has the highest consensus across all utterances class-wise. Statistically, I would expect this to be the MEAN similarity value.Then again, using MEAN could be bad as well, if the utterances for one routes are extremely similar and then for another one they are more heterogeneous. So for the current
HybridRouteLayer
, it seems evident that which method that is used is dependent on how the end user has defined the routes.Proposed solution
Allow developers to set aggregation themselves to allow them to find which solution works best.
This solution would be of relevant both for the
RouteLayer
andHybridRouteLayer
classes.EDIT: I can draft a PR to add this feature.
The text was updated successfully, but these errors were encountered: