Lightgbm model that predicts always majority class

Question

I have a Lightgbm Classifier multiclass with 3 classes.

With the following code, I fit the model:

over = ADASYN(sampling_strategy='not majority', random_state=7, n_jobs=-1)
under = RandomUnderSampler(sampling_strategy='majority', random_state=7)

X_resampled, y_resampled = over.fit_resample(X_train, y_train)
X_resampled, y_resampled = under.fit_resample(X_resampled, y_resampled)

clf = lightgbm.LGBMClassifier(**lgbm_params, nthreads=-1)
clf.class_weight = {0.0: 1/dfc.loc[0.0], 1.0: 1/dfc.loc[1.0], 2.0: 1/dfc.loc[2.0]}


clf.fit(X_resampled.drop('weight', axis=1), y_resampled)#, sample_weight=X_resampled['weight'])
y_pred = clf.predict(X_val.drop('weight', axis=1))
print('========== CLF ONLY ==========')
print(classification_report(y_pred, y_val))

When I run the fit method, this is my classification_report:

========== CLF ONLY ==========
              precision    recall  f1-score   support

         0.0       0.77      1.00      0.87     23130
         1.0       0.91      0.46      0.61      6774
         2.0       0.91      0.44      0.59      6769

    accuracy                           0.79     36673
   macro avg       0.86      0.63      0.69     36673
weighted avg       0.82      0.79      0.77     36673

However, with this code, even with the same data the classifier was trained on, the predicted class is always 0.0

    num = pd.DataFrame(MinMaxScaler().fit_transform(X)).iloc[-1:].copy()
    result = model.predict_proba(num, verbose=-1)

Why could this be happening? Thanks in advance