Error at getting performance metrics with multiclass classifier

[statscol]

Currently having an issue while trying to replicate one of the sample apps (Sample App scikit-learn - K-fold Cross Validation). Basically everything i did was loading IRIS dataset to Qlik following step by step instructions in the repo, however in the Train & test section i get the following error:

'Exception iterating responses: Target is multiclass but average='binary'. Please choose another average setting, one of [None, 'micro', 'macro', 'weighted'].'

This error comes out due to the scoring metric in GridSearchCV and when trying to compute performance scores like Precision or F1-score over a multiclass classification problem, how can i set the 'average' argument straight from Qlik?

[nabeel-oz]

Hi @statscol , you can specify the metric arguments when setting up the model. You will need to use the PyTools.sklearn_Setup_Adv function which allows for additional inputs. You can input empty string, i.e. '', for unnecessary parameters like DimReductionArgs.

// Set up a temporary table for the model parameters
[MODEL_INIT]:
LOAD
   [Model Name] as Model_Name,
   '' as EstimatorArgs,
   '' as ScalerArgs,
   'average=micro|str' as MetricArgs,
    '' as DimReductionArgs,
   'overwrite=true,cv=5,calculate_importances=true' as ExecutionArgs    
RESIDENT Estimators
WHERE [Model Name] = '$(vModel)';

[Result-Setup]:
LOAD
   model_name,
   result,
   timestamp
EXTENSION PyTools.sklearn_Setup_Adv(MODEL_INIT{Model_Name, EstimatorArgs, ScalerArgs, MetricArgs, DimReductionArgs, ExecutionArgs}); 

The input specifications for the metrics arguments are described here. If you just want to change the average parameter the argument would be average=micro|str or average=macro|str.

Unfortunately, there is a bit of an oversight in the code and you can't pass average=None at the moment as I thought that would always be the default. I will fix this in the next release.