3
$\begingroup$

I am trying to solve a multiclass classification problem using SVC as the base estimator and GridSearchCV to tune my results. Mentioned below is the code and the error being received:

svc_clf = SVC(C=0.7,tol=0.01,kernel='rbf',cache_size=500)
param_grid = {'C':np.linspace(0.1,1.0,10),'tol':[1e-7,1e-6,1e-5,1e-4,1e-3,1e-2,1e-1],'kernel':['rbf','poly'],'degree':[3,4,5,6,7]}

gs_svc = GridSearchCV(estimator=svc_clf,param_grid=param_grid,scoring='f1',cv=5)

gs_svc.fit(X_train,y_train)

Below is the error received:

ValueError: Target is multiclass but average='binary'. Please choose another average setting.

Upon my research I found that 'f1-score' isn't for multi-class classification. Please suggest which metric should be used for GridSearchCV for a multi-class classification problem ?

$\endgroup$
1
  • $\begingroup$ just use different scoring e.g. scoring='accuracy' $\endgroup$
    – quester
    Oct 11 '19 at 16:21
4
$\begingroup$

Accuracy might look tempting but not a good metric in general. In multilabel classification, for each class we'll have f1 score, precision, recall values etc. You need to decide how to average them, which is what the error is saying actually. The options are binary (which is the default one), micro, macro, weighted, samples. binary option needs positive and negative classes, and doesn't work in multilabel problems.

To reiterate sklearn documentation linked above, micro option calculates TP,FP etc. globally, while macro does it specific to each class and averages them. weighted is the weighted version of macro average that accounts for class imbalance.

And, this parameter needs to be passed into the scorer function, e.g.:

scorer = sklearn.metrics.make_scorer(sklearn.metrics.f1_score, average = 'weighted')
gs_svc = GridSearchCV(estimator=svc_clf,param_grid=param_grid,scoring=scorer,cv=5)
$\endgroup$
2
$\begingroup$

In addition to gunes' excellent answer, you may also use several scoring functions:

scoring = {'accuracy': make_scorer(accuracy_score),
           'precision': make_scorer(precision_score, average = 'macro'),
           'recall': make_scorer(recall_score, average = 'macro'),
           'f1_macro': make_scorer(f1_score, average = 'macro',
           'f1_weighted': make_scorer(f1_score, average = 'weighted')}

gs_svc = GridSearchCV(estimator=svc_clf,param_grid=param_grid,scoring=scoring,cv=5)
gs_svc.fit(X_train,y_train)
$\endgroup$
1
  • $\begingroup$ just to add info, this is pretty useful and works with GridSearchCV and RandomSearchCV. Though, you can´t do this if you are using BayesSearchCV - this was quite surprising, considering they all "follow the same interface" and belong to sklearn $\endgroup$ Jul 30 '20 at 14:10

This site is temporarily in read only mode and not accepting new answers.

Not the answer you're looking for? Browse other questions tagged .