Can't force class weights

1

I have a class imbalance where class are split like this:
0: 5.2%
1: 6.51%
2: 76.7%
3: 6.38%
4: 5.13%

I’m computing class weights with

from sklearn.utils.class_weight import compute_class_weight
class_weights = compute_class_weight(class_weight='balanced', classes=np.unique(ys), y=ys)
# produces : class_weights = [3.8412954, 3.06959094, 0.26054039, 3.13274541, 3.8984915]
...

loss_function = nn.CrossEntropyLoss(weight=class_weights)
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

However I still see that the model have bias towards predicting class 2 and not really learning anything other class.

image
image1885×444 32.3 KB

Any advice how to address this?

2

Did you increase the weights of the minority classes to “force” the model to learn them or did you only use the sklearn computed weights?

3

I just used the sklearn computed weights.

Did you increase the weights of the minority classes to “force” the model to learn them

I thought that by using the weights computed by sklearn I was doing just that.

4

Hey @ptrblck
I was wondering if you had some other suggestions here.
Thanks in advance.

5

Increase the weights of the minority classes until your model is overfitting these to make sure your code does not have any other issues.