Copilot based solution for some of the tasks.#2
Conversation
|
/gemini review |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request expands the machine learning lab script with new functions for anomaly detection, hyperparameter tuning, and ensemble methods. The review feedback identifies a critical issue in the LocalOutlierFactor implementation, which incorrectly uses fit_predict on test data and requires the novelty=True parameter to function correctly for model comparison. Additionally, the reviewer recommends refactoring the code to avoid redundant network requests for the same dataset and removing large blocks of commented-out code to improve maintainability.
| # Define the anomaly detection methods | ||
| methods = { | ||
| 'IsolationForest': ensemble.IsolationForest(contamination=0.1, random_state=42), | ||
| 'LocalOutlierFactor': neighbors.LocalOutlierFactor(n_neighbors=20), |
There was a problem hiding this comment.
To use LocalOutlierFactor for novelty detection (predicting on unseen data like the test set or the mesh grid), you must set novelty=True. Without this parameter, the model does not provide a predict method for new samples.
| 'LocalOutlierFactor': neighbors.LocalOutlierFactor(n_neighbors=20), | |
| 'LocalOutlierFactor': neighbors.LocalOutlierFactor(n_neighbors=20, novelty=True), |
| if method_name == 'LocalOutlierFactor': | ||
| method.fit(X_train_features) | ||
| y_pred_train = method.fit_predict(X_train_features) | ||
| y_pred_test = method.fit_predict(X_test_features) | ||
| else: | ||
| method.fit(X_train_features) | ||
| y_pred_train = method.predict(X_train_features) | ||
| y_pred_test = method.predict(X_test_features) | ||
|
|
||
| # Predict on mesh | ||
| if method_name == 'LocalOutlierFactor': | ||
| Z = method.fit_predict(np.c_[xx.ravel(), yy.ravel()]) | ||
| else: | ||
| Z = method.predict(np.c_[xx.ravel(), yy.ravel()]) |
There was a problem hiding this comment.
The current logic for LocalOutlierFactor is incorrect for a model comparison task. Calling fit_predict on the test set or the mesh grid ignores the training data and fits a new model on those specific samples. By setting novelty=True in the constructor, you can use the standard fit and predict methods consistently across all anomaly detection algorithms.
method.fit(X_train_features)
y_pred_test = method.predict(X_test_features)
# Predict on mesh
Z = method.predict(np.c_[xx.ravel(), yy.ravel()])|
|
||
| print_function_name() | ||
|
|
||
| X, y = datasets.fetch_openml('diabetes', as_frame=True, return_X_y=True) |
There was a problem hiding this comment.
The 'diabetes' dataset is fetched from OpenML multiple times across different functions (e.g., here and in todo_6). This is inefficient and makes the script dependent on a stable internet connection. Consider fetching the data once and passing it as an argument or using a caching mechanism.
| # print("\n\nLocalOutlierFactor with n_neighbors=20") | ||
| # lof = neighbors.LocalOutlierFactor(n_neighbors=20) | ||
| # y_pred_lof = lof.fit_predict(X_test_features) | ||
| # | ||
| # plt.figure(figsize=(10, 6)) | ||
| # plt.scatter( | ||
| # X_test_features[y_pred_lof == 1, 0], | ||
| # X_test_features[y_pred_lof == 1, 1], | ||
| # c='green', label='Inliers', alpha=0.6 | ||
| # ) | ||
| # plt.scatter( | ||
| # X_test_features[y_pred_lof == -1, 0], | ||
| # X_test_features[y_pred_lof == -1, 1], | ||
| # c='orange', label='Outliers', alpha=0.6 | ||
| # ) | ||
| # | ||
| # Create decision boundary for LOF using contour | ||
| # h = 0.5 | ||
| # x_min, x_max = X_test_features[:, 0].min() - 1, X_test_features[:, 0].max() + 1 | ||
| # y_min, y_max = X_test_features[:, 1].min() - 1, X_test_features[:, 1].max() + 1 | ||
| # | ||
| # xx, yy = np.meshgrid( | ||
| # np.arange(x_min, x_max, h), | ||
| # np.arange(y_min, y_max, h) | ||
| # ) | ||
| # | ||
| # Z_lof = lof.fit_predict(np.c_[xx.ravel(), yy.ravel()]) | ||
| # Z_lof = Z_lof.reshape(xx.shape) | ||
| # | ||
| # plt.contourf(xx, yy, Z_lof, alpha=0.3, levels=[-1, 0], colors=['orange']) | ||
| # plt.contourf(xx, yy, Z_lof, alpha=0.1, levels=[0, 1], colors=['green']) | ||
| # | ||
| # plt.xlabel('plas') | ||
| # plt.ylabel('mass') | ||
| # plt.title('LocalOutlierFactor (n_neighbors=20)') | ||
| # plt.legend() |
There was a problem hiding this comment.
This large block of commented-out code should be removed to improve readability. If these experiments are intended to be optional or for future reference, consider moving them to a separate script or a dedicated function. Note that the code also contains a bug in the LocalOutlierFactor usage (missing novelty=True).
1 participant
No description provided.