Features/search

[VanAnh-13]

Refactor Config và Metric

Config: Áp dụng cơ chế đọc cấu hình từ YAML cho các chiến lược search (có hỗ trợ fallback).

Scoring: Sửa đổi engine để xử lý động các metric tùy chỉnh và hỗ trợ nhiều phương pháp tính trung bình (averaging methods).

Code quality: Dịch comment sang tiếng Việt để tăng tính dễ hiểu, chuẩn hóa các hàm tiện ích và bổ sung logger.

[Copilot]

Pull request overview

This PR refactors the AutoML search strategy configuration and scoring system to improve flexibility and maintainability. The changes introduce YAML-based configuration management with fallback support, enable dynamic metric handling with both macro and weighted averaging methods, and translate comments to Vietnamese for better clarity.

Key Changes:

• Implemented centralized YAML configuration loading with a two-tier fallback system (primary config → default config) across all search strategies
• Modified scoring engine to generate both macro and weighted variants for precision, recall, and f1 metrics, providing comprehensive evaluation for imbalanced datasets
• Extracted and centralized numpy type conversion and cache cleanup logic into reusable helper methods

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
base.py	Added _load_yaml_config() method for centralized config loading with fallback, refactored get_default_config() to read from YAML files
base_config.yml	Primary base configuration file for all search strategies with CV, scoring, logging, and parallel processing settings
base_default_config.yml	Fallback base configuration with identical structure to base_config.yml
grid_search.py	Refactored to use centralized YAML config loading, enhanced _get_params_hash() to include model base parameters for better cache uniqueness
grid_search_config.yml	Grid search specific configuration for dispatch, training, parallel evaluation, and caching
grid_search_default_config.yml	Fallback grid search configuration with identical structure
genetic_algorithm.py	Replaced custom config loading with base class method, moved numpy type conversion to centralized helper, extracted generation creation logic into _create_next_generation() method, simplified metric handling to work with pre-configured scoring dict from engine.py
genetic_algorithm_config.yml	GA configuration covering population, generation, genetic operators, performance optimization, and adaptive features
genetic_algorithm_default_config.yml	Fallback GA configuration with identical structure
bayesian_search.py	Refactored to use centralized YAML config loading, simplified metric handling to accept pre-configured scoring from engine.py, uses _finalize_results() for consistent cleanup
bayesian_search_config.yml	Bayesian search configuration for optimization parameters, acquisition function, and early stopping
bayesian_search_default_config.yml	Fallback Bayesian search configuration
engine.py	Modified to create both macro and weighted scorers for precision, recall, f1 metrics; updated score extraction logic to retrieve both averaging variants; translated comments to Vietnamese

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

[nitpick] The code temporarily modifies self.config['crossover_rate'] at lines 388-391 to use an adaptive rate, which is a side effect that could cause issues if the code is interrupted or if other code reads this config value during execution. Consider passing the crossover rate as a parameter to _crossover instead of modifying the shared config state.

For example:

child1, child2 = self._crossover(parent1, parent2, crossover_rate=adaptive_crossover_rate)

This would make the code more maintainable and avoid potential race conditions in concurrent scenarios.

[Copilot]

The metric_sort parameter is used to look up rank_test_{metric_sort} in cv_results at line 185. However, when metrics like 'precision', 'recall', or 'f1' are processed, they are split into '_macro' and '_weighted' variants (lines 129-137). This means if metric_sort is 'precision', the lookup for 'rank_test_precision' will fail since only 'rank_test_precision_macro' and 'rank_test_precision_weighted' exist in cv_results.

Consider either:

1. Transforming metric_sort to include the suffix (e.g., 'precision' -> 'precision_macro') before passing to the search strategy, or
2. Updating the lookup logic to handle both base names and suffixed names with a fallback mechanism.

Suggested change

	rank_key = f'rank_test_{metric_sort}'
	if rank_key in cv_results:
	rank_array = np.array(cv_results[rank_key])
	else:
	# Handle suffixed metric names for precision, recall, f1
	rank_array = None
	if metric_sort in ['precision', 'recall', 'f1']:
	# Prefer macro, fallback to weighted
	for avg_type in ['macro', 'weighted']:
	rank_key = f'rank_test_{metric_sort}_{avg_type}'
	if rank_key in cv_results:
	rank_array = np.array(cv_results[rank_key])
	break
	else:
	rank_key = f'rank_test_{metric_sort}'
	if rank_key in cv_results:
	rank_array = np.array(cv_results[rank_key])
	if rank_array is None:

[Copilot]

In the fallback case where best_all_scores is created from scratch (line 751), the code attempts to set the primary metric score at line 754: if primary_metric in best_all_scores. However, if metric_sort is a base metric name like 'precision' but the scoring config creates suffixed metrics like 'precision_macro' and 'precision_weighted', this condition will always be False. The best_score will not be assigned to the correct metric.

Consider checking for the presence of the primary metric with suffixes (e.g., if primary_metric is 'precision', check for 'precision_macro' or 'precision_weighted') or ensuring metric_sort is transformed to match the actual scoring keys.

Suggested change

	best_all_scores[primary_metric] = best_score
	best_all_scores[primary_metric] = best_score
	else:
	# Tìm key bắt đầu bằng primary_metric (ví dụ: 'precision' -> 'precision_macro')
	matching_keys = [k for k in best_all_scores if k.startswith(primary_metric)]
	if matching_keys:
	best_all_scores[matching_keys[0]] = best_score

[Copilot]

The primary_metric retrieved from config may not match the actual metric keys in the scores dictionary. For example, if metric_sort is 'precision', but engine.py creates scoring with keys 'precision_macro' and 'precision_weighted', the lookup scores.get(primary_metric, 0.0) will return the default value 0.0 instead of the actual score. This would cause incorrect fitness evaluation and optimization.

Consider either:

1. Ensuring metric_sort is transformed to include the suffix before being passed to the search strategy, or
2. Implementing a fallback lookup that tries the base name first, then checks for '_macro' and '_weighted' suffixes.

[Copilot]

Similar to the issue in genetic_algorithm.py and engine.py, the primary_metric from metric_sort config may not match the actual keys in the scoring results. If metric_sort is 'precision' but the scoring dict has 'precision_macro' and 'precision_weighted', the lookup at line 204 objective.last_metrics.get(primary_metric, ...) will return the default value instead of the actual metric score. This would cause incorrect optimization.

Consider implementing a fallback mechanism that handles both base metric names and their suffixed variants, or ensuring metric_sort is transformed before being passed to the search strategy.

Suggested change

	score = objective.last_metrics.get(primary_metric, objective.last_metrics.get('accuracy', 0.0))
	score = objective.last_metrics.get(primary_metric)
	if score is None:
	# Fallback: try to find a key that starts with the primary_metric
	for k in objective.last_metrics:
	if k == primary_metric or k.startswith(primary_metric + "_"):
	score = objective.last_metrics[k]
	break
	if score is None:
	score = objective.last_metrics.get('accuracy', 0.0)

[optivisionlab]

OKE