Combining several imperfect classifiers to reduce overall classification error.ipynb

Combining several imperfect classifiers to reduce overall classification error.ipynb

grid_search.py

from sklearn.svm import SVC
from sklearn.ensemble import VotingClassifier
from sklearn.metrics import accuracy_score
from sklearn.model_selection import ParameterGrid


# Define the search space
subset_seed_range = range(0, 100)  # Range of seeds to try
param_grid = {
    'C': [0.1, 1, 10],
    'gamma': [0.001, 0.01, 0.1, 1],
    'kernel': ['rbf'],
    'random_state': range(0,5)
}
all_combinations = [(seed, params) for seed in subset_seed_range for params in ParameterGrid(param_grid)]

# Initialize variables to track the best performance
best_ensemble_accuracy = 0
best_combination = None

n_passed = 0

# Perform the search
for seed, params in all_combinations:
    #print(seed)
    #print(params)
    # Create the subsets using the seed
    np.random.seed(seed)
    subset_indices_1 = np.random.choice(range(len(X)), size=int(len(X)/3), replace=False)
    subset_indices_2 = np.random.choice(list(set(range(len(X))) - set(subset_indices_1)), size=int(len(X)/3), replace=False)
    subset_indices_3 = list(set(range(len(X))) - set(subset_indices_1) - set(subset_indices_2))

    X_sub1, y_sub1 = X[subset_indices_1], y[subset_indices_1]
    X_sub2, y_sub2 = X[subset_indices_2], y[subset_indices_2]
    X_sub3, y_sub3 = X[subset_indices_3], y[subset_indices_3]
    
    # Train SVC models on each subset with the given parameters
    classifiers = []
    for i, (X_sub, y_sub) in enumerate([(X_sub1, y_sub1), (X_sub2, y_sub2), (X_sub3, y_sub3)]):
        clf = SVC(**params).fit(X_sub, y_sub)
        classifiers.append(clf)
    
    # Create the ensemble
    ensemble = VotingClassifier(estimators=[(f'clf{i}', clf) for i, clf in enumerate(classifiers)], voting='hard')
    ensemble.fit(X, y)
    
    # Evaluate the ensemble and base classifiers
    ensemble_accuracy = accuracy_score(y, ensemble.predict(X))
    base_accuracies = [accuracy_score(y, clf.predict(X)) for clf in classifiers]
    
    # We don't want 'bad' examples
    passed = all(acc < ensemble_accuracy for acc in base_accuracies)
    
    if passed:
        n_passed = n_passed + 1
    
    # Check if this combination is the best so far
    if ensemble_accuracy > best_ensemble_accuracy and passed:
        best_ensemble_accuracy = ensemble_accuracy
        best_combination = (seed, params)
    
    #print("----")

# Output the results
print(f"Best ensemble accuracy: {best_ensemble_accuracy:.2f}")
print(f"Best combination: Seed={best_combination[0]}, Parameters={best_combination[1]}")

Združevanje_več_klasifikatorjev (alt).ipynb

Združevanje_več_klasifikatorjev.ipynb