from sklearn.linear_model import ElasticNetCV # Regression
from sklearn.linear_model import RidgeClassifierCV # Classification
from sklearn.metrics import (
# Regression metrics
median_absolute_error,
r2_score,
# Classification metrics
accuracy_score,
precision_score,
recall_score,
f1_score,
roc_auc_score,
plot_confusion_matrix,
plot_roc_curve,
plot_precision_recall_curve,
)
if not os.path.exists(
"../results/regression_vectorisers_results.csv"
) or not os.path.exists("../results/classification_vectorisers_results.csv"):
results_reg = []
results_cls = []
for vectorizer_name, vectorizer in vectorizers.items():
# Vectorization of the corpus
X_train_vec = vectorizer.fit_transform(X_train)
X_test_vec = vectorizer.transform(X_test)
# Tokens distribution
words_count = pd.Series(
X_train_vec.sum(axis=0).tolist()[0],
index=vectorizer.get_feature_names_out(),
)
# Top 20 tokens by occurrences
top_20_count = words_count.sort_values(ascending=False).head(20)
fig = px.bar(
top_20_count,
x=top_20_count.index,
y=top_20_count.values,
labels={"x": "Word", "y": "Count"},
title=f"{vectorizer_name} : Top 20 frequent words in reviews (vocabulary = {len(words_count)} words)",
color=top_20_count.values,
)
fig.show()
# Regression model
print("Regression")
# Train the regression model
reg = ElasticNetCV(random_state=42, n_jobs=-1).fit(X_train_vec, y_train)
# Comute the coefficients
coefs_reg = pd.Series(reg.coef_, index=vectorizer.get_feature_names_out())
# Top 20 tokens by importance (positive and negative)
top_20_coefs_reg = (
coefs_reg.nlargest(10).append(coefs_reg.nsmallest(10)).sort_values()
)
fig = px.bar(
top_20_coefs_reg,
x=top_20_coefs_reg.index,
y=top_20_coefs_reg.values,
labels={"x": "Word", "y": "Count"},
title=f"{vectorizer_name} : Top 20 important words in reviews (vocabulary = {len(words_count)} words)",
color=top_20_coefs_reg.values,
)
fig.show()
# Test the regression model
y_pred_reg = reg.predict(X_test_vec)
fig = px.box(
x=y_test,
y=y_pred_reg,
labels={"x": "Actual", "y": "Predicted"},
title=f"{vectorizer_name} : Actual vs Predicted / R² = {round(r2_score(y_test, y_pred_reg), 3)} / MAE = {round(median_absolute_error(y_test, y_pred_reg), 3)}",
color=y_test,
)
fig.show()
# Keep the results
results_reg.append(
{
"vectorizer": vectorizer_name,
"vocabulary_size": len(words_count),
"r2_score": r2_score(y_test, y_pred_reg),
"median_absolute_error": median_absolute_error(y_test, y_pred_reg),
}
)
print()
print(f"{vectorizer_name}")
print(
f"vocabulary = {len(words_count)} words / R² = {round(r2_score(y_test, y_pred_reg), 3)} / MAE = {round(median_absolute_error(y_test, y_pred_reg), 3)}"
)
print()
# Classification model
print("Classification")
# Train the regression model
cls = RidgeClassifierCV().fit(X_train_vec, y_train_bi)
# Comute the coefficients
coefs_cls = pd.Series(
cls.coef_[0],
index=vectorizer.get_feature_names_out(),
)
# Top 20 tokens by importance (positive and negative)
top_20_coefs_cls = (
coefs_cls.nlargest(10).append(coefs_cls.nsmallest(10)).sort_values()
)
fig = px.bar(
top_20_coefs_cls,
x=top_20_coefs_cls.index,
y=top_20_coefs_cls.values,
labels={"x": "Word", "y": "Count"},
title=f"{vectorizer_name} : Top 20 important words in reviews (vocabulary = {len(words_count)} words)",
color=top_20_coefs_cls.values,
)
fig.show()
# Test the regression model
y_pred_cls = cls.predict(X_test_vec)
plot_confusion_matrix(
estimator=cls,
X=X_test_vec,
y_true=y_test_bi,
)
plt.show()
plot_roc_curve(
estimator=cls,
X=X_test_vec,
y=y_test_bi,
)
plt.show()
plot_precision_recall_curve(
estimator=cls,
X=X_test_vec,
y=y_test_bi,
)
plt.show()
# Keep the results
results_cls.append(
{
"vectorizer": vectorizer_name,
"vocabulary_size": len(words_count),
"accuracy_score": accuracy_score(y_test_bi, y_pred_cls),
"precision_score": precision_score(y_test_bi, y_pred_cls),
"recall_score": recall_score(y_test_bi, y_pred_cls),
"f1_score": f1_score(y_test_bi, y_pred_cls),
"roc_auc_score": roc_auc_score(y_test_bi, y_pred_cls),
}
)
print()
print(f"{vectorizer_name}")
print(
f"vocabulary = {len(words_count)} words / accuracy_score = {round(accuracy_score(y_test_bi, y_pred_cls), 3)} / precision_score = {round(precision_score(y_test_bi, y_pred_cls), 3)} / recall_score = {round(recall_score(y_test_bi, y_pred_cls), 3)} / f1_score = {round(f1_score(y_test_bi, y_pred_cls), 3)} / roc_auc_score = {round(roc_auc_score(y_test_bi, y_pred_cls), 3)}"
)
print()
print(
pd.DataFrame(results_reg).sort_values(
by=["r2_score", "median_absolute_error"],
ascending=[False, True],
)
)
with open("../results/regression_vectorisers_results.csv", "w") as f:
f.write(
pd.DataFrame(results_reg)
.sort_values(
by=["r2_score", "median_absolute_error"],
ascending=[False, True],
)
.to_csv(index=False)
)
print(
pd.DataFrame(results_cls).sort_values(
by=["roc_auc_score", "f1_score"],
ascending=[False, False],
)
)
with open("../results/classification_vectorisers_results.csv", "w") as f:
f.write(
pd.DataFrame(results_cls)
.sort_values(
by=["roc_auc_score", "f1_score"],
ascending=[False, False],
)
.to_csv(index=False)
)
else:
results_reg_df = pd.read_csv("../results/regression_vectorisers_results.csv")
print(
results_reg_df.sort_values(
by=["r2_score", "median_absolute_error"],
ascending=[False, True],
)
)
results_cls_df = pd.read_csv("../results/classification_vectorisers_results.csv")
print(
results_cls_df.sort_values(
by=["roc_auc_score", "f1_score"],
ascending=[False, False],
)
)
