def best_ranker_accuracy(X,
y,
evaluator,
min_features=None,
max_features=None):
max_feat = max_features or len(X.columns)
min_feat = min_features or 1
log.debug(
f"Calculating accuracy of {min_feat} to {max_feat} ranked columns")
best = None
for feature_count in range(min_feat, max_feat):
n_scores, selected_columns = evaluator(X, y, feature_count)
mean_score = np.mean(n_scores)
if not best or mean_score > best[0]:
best = (np.mean(n_scores), selected_columns)
return best
def evaluate_by_rank(X, y, n_features):
ranker = RFE(DecisionTreeClassifier(), n_features_to_select=n_features)
ranks = ranker.fit(X, y)
selected_columns = list(X.columns[ranks.support_])
pipeline = Pipeline(steps=[('s', ranker), ('m', DecisionTreeClassifier())])
# evaluate model
cv = RepeatedStratifiedKFold(n_splits=5, n_repeats=3, random_state=1)
n_scores = cross_val_score(pipeline,
X,
y,
scoring='accuracy',
cv=cv,
n_jobs=-1,
error_score='raise')
# report performance
log.debug(
f"Accuracy: {np.mean(n_scores):.3f} ({np.std(n_scores):.3f}): COLS {selected_columns}"
)
return n_scores, selected_columns
def decision_tree(X, y, importance_threshold=0.07):
classer = DecisionTreeClassifier()
log.debug(f"churn columns: {X.columns}")
result = classer.fit(X, y)
log.debug(f"CLASSER RESULT: {result}")
important_cols = [
c for i, c in enumerate(X.columns)
if result.feature_importances_[i] > importance_threshold
]
log.debug(f"CLASSER IMPORTANT COLUMNS: {important_cols}")
def evaluate_by_pca(X, y, n_features):
pca = PCA(n_components=n_features)
pca.fit(X)
log.debug(pca.explained_variance_ratio_)
log.debug(pca.singular_values_)
return pca.score(X, y), 0