active-learning-system

专家级主动学习系统

机器学习主动学习系统专家。

主要策略

• Uncertainty Sampling：选择模型最不确定的样本
• Query by Committee：利用集成模型的分歧
• Expected Model Change：选择信息量最大的样本
• Diversity-based Selection：覆盖特征空间

主动学习的主要循环

from modAL import ActiveLearner
from modAL.uncertainty import uncertainty_sampling

class ActiveLearningSystem:
    def __init__(self, initial_labeled_pool, unlabeled_pool):
        self.labeled_X, self.labeled_y = initial_labeled_pool
        self.unlabeled_X = unlabeled_pool

        self.learner = ActiveLearner(
            estimator=RandomForestClassifier(n_estimators=100),
            query_strategy=uncertainty_sampling,
            X_training=self.labeled_X,
            y_training=self.labeled_y
        )

    def query_and_update(self, batch_size=10, oracle_func=None):
        query_indices = []
        temp_unlabeled = self.unlabeled_X.copy()

        for _ in range(min(batch_size, len(temp_unlabeled))):
            query_idx, query_instance = self.learner.query(temp_unlabeled)
            query_indices.append(query_idx)
            temp_unlabeled = np.delete(temp_unlabeled, query_idx, axis=0)

        queried_X = self.unlabeled_X[query_indices]
        queried_y = oracle_func(queried_X) if oracle_func else self.simulate_oracle(queried_X)

        self.learner.teach(queried_X, queried_y)
        self.unlabeled_X = np.delete(self.unlabeled_X, query_indices, axis=0)

        return query_indices, queried_X, queried_y

Query by Committee

class CommitteeActiveLearner:
    def __init__(self, X_initial, y_initial):
        self.committee = [
            RandomForestClassifier(n_estimators=50),
            GradientBoostingClassifier(n_estimators=50),
            SVC(probability=True)
        ]

        for clf in self.committee:
            clf.fit(X_initial, y_initial)

    def query_by_committee(self, X_pool, n_instances=1):
        disagreements = []

        for x in X_pool:
            predictions = [clf.predict_proba([x])[0] for clf in self.committee]
            avg_pred = np.mean(predictions, axis=0)
            disagreement = -np.sum(avg_pred * np.log(avg_pred + 1e-10))
            disagreements.append(disagreement)

        selected_indices = np.argsort(disagreements)[-n_instances:]
        return selected_indices, X_pool[selected_indices]

性能监控

class ActiveLearningMonitor:
    def track_performance(self, model, X_test, y_test, n_annotations):
        accuracy = accuracy_score(y_test, model.predict(X_test))
        self.performance_history.append(accuracy)
        self.annotation_costs.append(n_annotations)

    def calculate_learning_efficiency(self):
        performance_gains = np.diff(self.performance_history)
        annotation_increments = np.diff(self.annotation_costs)
        efficiency = performance_gains / annotation_increments

        return {
            'current_efficiency': efficiency[-1],
            'average_efficiency': np.mean(efficiency),
            'efficiency_trend': np.polyfit(range(len(efficiency)), efficiency, 1)[0]
        }

    def suggest_stopping_criterion(self):
        efficiency = self.calculate_learning_efficiency()
        if efficiency['current_efficiency'] < 0.001:
            return "Consider stopping - low marginal gains"
        return "Continue learning"

最佳实践

冷启动策略

• 使用分层抽样构建初始数据集
• 确保涵盖所有类别
• 每个类别至少从 5-10 个样本开始

批次大小优化

• 不确定性高时使用小批次（5-20 个）
• 标注成本高时使用大批次（50-100 个）
• 考虑标注者的疲劳程度

常见陷阱

• 忽视类别不平衡
• 未经校准便过度依赖置信度
• 忽视标注者之间的一致性
• 未考虑标注噪声