Ensemble Example

This example works on the Phishing dataset. The goal is to predict whether a sample is phishing or not (binary classification) using an ensemble method. We will use AdaBoost with a Hoeffding Tree base classifier.

Models

We will use the following models and components:

• HoeffdingTreeClassifier as the base estimator for AdaBoost.
• AdaBoostClassifier as the ensemble method.

We will also use the following metric for evaluation:

• LogLoss

Preprocessing

No explicit preprocessing is defined in this example, as the Hoeffding Tree can handle raw features directly. If your data requires preprocessing, you can add preprocessors as needed.

Beaver File Structure

Let's see how we would write the beaver file:

Connector

We start by defining the connector, specifying the Kafka bootstrap servers and security protocol.

connector {
        bootstrap_servers = "localhost:39092"
        security_protocol = "plaintext"
        consumer_group = 'ensemble'
        auto_offset_reset = "earliest"
}

Models

We define the base classifier and the AdaBoost ensemble:

algorithm <HoeffdingTreeClassifier> hoeff
    params:
        split_criterion='gini',
        delta=1e-5,
        grace_period=2000

algorithm <AdaBoostClassifier> adaboost
    params:
        model=hoeff,
        n_models=5,
        seed=42

Metric

We define the evaluation metric:

metric <LogLoss> logloss

Data

We define the data source and specify the target feature:

data Phishing {
    input_topic = "Phishing"
    features:
        target_feature = is_phishing
}

Pipeline

Finally, we define the pipeline that brings everything together:

pipeline adaboostPipeline {
    output_topic = 'adaboostPipeline'
    data = Phishing
    algorithm = adaboost
    metrics = logloss
}

And that's it! This setup allows you to use AdaBoost with Hoeffding Trees to improve classification performance on the Phishing dataset by

connector {
        bootstrap_servers = "localhost:39092"
        security_protocol = "plaintext"
        consumer_group = 'ensemble'
        auto_offset_reset = "earliest"
}

algorithm <HoeffdingTreeClassifier> hoeff
    params:
        split_criterion='gini',
        delta=1e-5,
        grace_period=2000



algorithm <AdaBoostClassifier> adaboost
    params:
        model=hoeff,
        n_models=5,
        seed=42



metric <LogLoss> logloss

data Phishing {

    input_topic = "Phishing"
    features:
        target_feature = is_phishing
}

pipeline adaboostPipeline {
    output_topic = 'adaboostPipeline'
    data = Phishing
    algorithm = adaboost
    metrics = logloss

}