balancr 0.1.1

A unified framework for analysing and comparing techniques for handling imbalanced datasets

Balancr: A Unified Framework for Analysing Data Balancing Techniques

A comprehensive framework and CLI tool for analysing and comparing different techniques for handling imbalanced datasets in machine learning. Balancr makes it easier to compare balancing algorithims against a wide range of classifers

Overview

Imbalanced datasets are a significant challenge in machine learning, particularly in areas such as:

• Medical diagnosis
• Fraud detection
• Network intrusion detection
• Rare event prediction

Balancr allows you to:

• Compare different balancing techniques (e.g., SMOTE, ADASYN, random undersampling), and the same technqiues with different configurations, against multiple classifiers
• Evaluate performance using relevant metrics
• Visualise results and class distributions
• Generate balanced datasets using various methods
• Customise the evaluation process with different classifiers

Features

Core Functionality:

• CLI Interface: Simple command-line interface for full workflow
• Data Loading: Support for CSV, and provides a data quality check
• Preprocessing: Configurable preprocessing functionality including handling data quality issues, scaling, and encoding categorical features
• Dynamic Technique Discovery: Automatic discovery of techniques from imbalanced-learn
• Custom Technique Registration: Register your own balancing techniques
• Classifier Selection: Compare performance across multiple classifiers. Fine-tune parameters via the configuration file
• Custom Classifier Registration: Register your own classifier implementations
• Comprehensive Metric Evaluation: Get metrics specific to imbalanced learning
• Visualisation Suite: Plots for class distributions, metrics comparison, and learning curves (more to come)
• Flexible Configuration: Configure every aspect via CLI or configuration file

Available Metrics

• Accuracy
• Precision
• Recall
• F1-score
• ROC AUC
• G-mean
• Specificity
• Cross-validation scores

Visualisations

• Class distribution comparisons
• Performance metric comparisons
• Learning curves
• Results comparison plots

Installation

# From PyPI (recommended)
pip install balancr

# From source
git clone https://gitlab.eeecs.qub.ac.uk/40353634/csc3002-balancing-techniques-framework.git
cd balancing-techniques-framework
pip install -e .

Command-Line Interface

Balancr provides a comprehensive CLI to help you analyse imbalanced datasets:

  ____        _                       
 | __ )  __ _| | __ _ _ __   ___ _ __ 
 |  _ \ / _` | |/ _` | '_ \ / __| '__|
 | |_) | (_| | | (_| | | | | (__| |   
 |____/ \__,_|_|\__,_|_| |_|\\___|_|   
                                     

Quick Start - CLI

Here's a complete workflow using the CLI:

# Load your dataset
balancr load-data dataset.csv -t target_column

# Configure preprocessing
balancr preprocess --scale standard --handle-missing mean --encode auto

# Select balancing techniques to compare
balancr select-techniques SMOTE RandomUnderSampler ADASYN

# Select classifiers for evaluation
balancr select-classifiers RandomForestClassifier LogisticRegression

# Configure metrics
balancr configure-metrics --metrics precision recall f1 roc_auc

# Configure visualisations
balancr configure-visualisations --types all --save-formats png

# Configure evaluation settings
balancr configure-evaluation --test-size 0.3 --cross-validation 5

# Run the comparison
balancr run --output-dir results/experiment1

Available Commands

Command	Description	Example
load-data	Load a dataset for analysis	balancr load-data dataset.csv -t target
preprocess	Configure preprocessing options	balancr preprocess --scale standard --handle-missing mean
select-techniques	Select balancing techniques	balancr select-techniques SMOTE ADASYN
register-techniques	Register custom techniques	balancr register-techniques my_technique.py
select-classifiers	Select classifiers for evaluation	balancr select-classifiers RandomForestClassifier
register-classifiers	Register custom classifiers	balancr register-classifiers my_classifier.py
configure-metrics	Configure evaluation metrics	balancr configure-metrics --metrics precision recall f1
configure-visualisations	Configure visualisation options	balancr configure-visualisations --types all
configure-evaluation	Configure model evaluation settings	balancr configure-evaluation --test-size 0.3
run	Run comparison of techniques	balancr run --output-dir results
reset	Reset configuration to defaults	balancr reset

Python API

Balancr can also be used as a Python library:

from balancr.imbalance_analyser import BalancingFramework

# Initialize the framework
framework = BalancingFramework()

# Load your dataset
framework.load_data(
    file_path="path/to/your/data.csv",
    target_column="target",
    feature_columns=["feature1", "feature2", "feature3"]
)

# Preprocess the data
framework.preprocess_data(
    handle_missing="mean",
    scale="standard",
    encode="auto"
)

# Apply balancing techniques
balanced_datasets = framework.apply_balancing_techniques(
    technique_names=["SMOTE", "RandomUnderSampler", "ADASYN"],
    test_size=0.2
)

# Train and evaluate classifiers
results = framework.train_classifiers(
    classifier_configs={
        "RandomForestClassifier": {"n_estimators": 100, "random_state": 42},
        "LogisticRegression": {"C": 1.0, "random_state": 42}
    },
    enable_cv=True,
    cv_folds=5
)

# Generate visualisations
framework.compare_balanced_class_distributions(
    save_path="results/class_distributions.png"
)

# Generate learning curves
framework.generate_learning_curves(
    classifier_name="RandomForestClassifier",
    save_path="results/learning_curves.png"
)

# Save results
framework.save_classifier_results(
    "results/metrics_results.csv",
    classifier_name="RandomForestClassifier"
)

Creating Custom Techniques

You can create and register your own balancing techniques:

from balancr.base import BaseBalancer
import numpy as np

class MyCustomBalancer(BaseBalancer):
    """A custom balancing technique that implements your logic"""
    
    def balance(self, X: np.ndarray, y: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
        # Implement your balancing logic here
        # This should return the balanced X and y
        
        return X_balanced, y_balanced

Register your technique using the CLI:

balancr register-techniques my_custom_technique.py

Or using the Python API:

from balancr.technique_registry import TechniqueRegistry
from my_custom_technique import MyCustomBalancer

registry = TechniqueRegistry()
registry.register_custom_technique("MyCustomBalancer", MyCustomBalancer)

Creating Custom Classifiers

You can create and register your own classifiers:

from sklearn.base import BaseEstimator
import numpy as np

class CustomClassifier(BaseEstimator):
    def __init__(self, n_estimators, random_state):
        self.n_estimators = n_estimators
        self.random_state = random_state

    def fit(self, X, y):
        # Implement your training logic here
        # Return self. Fitted estimator.
        return self

    def predict(self, X):
        # Implement your prediction logic here
        # Return your predictions/list of predicitons
        return np.zeros(len(X))

Register your classifier using the CLI:

balancr register-classifier my_custom_classifier.py

Or using the Python API:

from balancr.classifier_registry import ClassifierRegistry
from my_custom_classifier import MyCustomClassifier

registry = ClassifierRegistry()
registry.register_custom_classifier("MyCustomClassifier", MyCustomClassifier)

Extra Configuration Tips

Manual Configurations

For more control, all confiurgations are stored in balancr's config file (default location: ~/.balancr/config.json)

Comparing Balancers/Classifiers Against Themselves

To be able to compare a balancing technique or classifier against itself, but with different parameters, as is often the case, extra configuration is required

To compare a balancing technique against itself:

First, select the balancer you want to compare:

balancr select-techniques SMOTE

Then in balancr's config file (default location: ~/.balancr/config.json), you should see the config settings of your selected technique:

  "balancing_techniques": {
    "SMOTE": {
      "sampling_strategy": "auto",
      "random_state": 42,
      "k_neighbors": 3,
      "n_jobs": null
    }
  },

Create a copy of this technique config, and make sure to change the name to contain a valid suffix (suffix needs to start with _ or -), e.g. SMOTE_v2, SMOTE-2, SMOTE_ChangedParams, etc.:

  "balancing_techniques": {
    "SMOTE": {
      "sampling_strategy": "auto",
      "random_state": 42,
      "k_neighbors": 3,
      "n_jobs": null
    },
    "SMOTE_v2": {
      "sampling_strategy": "auto",
      "random_state": 42,
      "k_neighbors": 3,
      "n_jobs": null
    }
  },

You can then change the desired parameters. In this exmaple, we change k_neighbors from 3 to 5:

  "balancing_techniques": {
    "SMOTE": {
      "sampling_strategy": "auto",
      "random_state": 42,
      "k_neighbors": 3,
      "n_jobs": null
    },
    "SMOTE_v2": {
      "sampling_strategy": "auto",
      "random_state": 42,
      "k_neighbors": 5, 
      "n_jobs": null
    }
  },

To compare a classifier against itself:

First, select the classifier you want to compare:

balancr select-classifiers RandomForestClassifier

Then in balancr's config file (default location: ~/.balancr/config.json), you should see the config settings of your selected classifier:

    "classifiers": {
    "RandomForestClassifier": {
      "n_estimators": 100,
      "criterion": "gini",
      "max_depth": null,
      "min_samples_split": 2,
      "min_samples_leaf": 1,
      "min_weight_fraction_leaf": 0.0,
      "max_features": "sqrt",
      "max_leaf_nodes": null,
      "min_impurity_decrease": 0.0,
      "bootstrap": true,
      "oob_score": false,
      "n_jobs": null,
      "random_state": null,
      "verbose": 0,
      "warm_start": false,
      "class_weight": null,
      "ccp_alpha": 0.0,
      "max_samples": null,
      "monotonic_cst": null
    }
  },

Create a copy of this classifier config, and make sure to change the name to contain a valid suffix (suffix needs to start with _ or -), e.g. RandomForestClassifier_v2, RandomForestClassifier-2, RandomForestClassifier_ChangedParams, etc.:

    "classifiers": {
    "RandomForestClassifier": {
      "n_estimators": 100,
      "criterion": "gini",
      "max_depth": null,
      "min_samples_split": 2,
      "min_samples_leaf": 1,
      "min_weight_fraction_leaf": 0.0,
      "max_features": "sqrt",
      "max_leaf_nodes": null,
      "min_impurity_decrease": 0.0,
      "bootstrap": true,
      "oob_score": false,
      "n_jobs": null,
      "random_state": null,
      "verbose": 0,
      "warm_start": false,
      "class_weight": null,
      "ccp_alpha": 0.0,
      "max_samples": null,
      "monotonic_cst": null
    },
    "RandomForestClassifier_v2": {
      "n_estimators": 100,
      "criterion": "gini",
      "max_depth": null,
      "min_samples_split": 2,
      "min_samples_leaf": 1,
      "min_weight_fraction_leaf": 0.0,
      "max_features": "sqrt",
      "max_leaf_nodes": null,
      "min_impurity_decrease": 0.0,
      "bootstrap": true,
      "oob_score": false,
      "n_jobs": null,
      "random_state": null,
      "verbose": 0,
      "warm_start": false,
      "class_weight": null,
      "ccp_alpha": 0.0,
      "max_samples": null,
      "monotonic_cst": null
    }
  },

You can then change the desired parameters. In this exmaple, we change n_estimators from 100 to 200:

    "classifiers": {
    "RandomForestClassifier": {
      "n_estimators": 100,
      "criterion": "gini",
      "max_depth": null,
      "min_samples_split": 2,
      "min_samples_leaf": 1,
      "min_weight_fraction_leaf": 0.0,
      "max_features": "sqrt",
      "max_leaf_nodes": null,
      "min_impurity_decrease": 0.0,
      "bootstrap": true,
      "oob_score": false,
      "n_jobs": null,
      "random_state": null,
      "verbose": 0,
      "warm_start": false,
      "class_weight": null,
      "ccp_alpha": 0.0,
      "max_samples": null,
      "monotonic_cst": null
    },
    "RandomForestClassifier_v2": {
      "n_estimators": 200,
      "criterion": "gini",
      "max_depth": null,
      "min_samples_split": 2,
      "min_samples_leaf": 1,
      "min_weight_fraction_leaf": 0.0,
      "max_features": "sqrt",
      "max_leaf_nodes": null,
      "min_impurity_decrease": 0.0,
      "bootstrap": true,
      "oob_score": false,
      "n_jobs": null,
      "random_state": null,
      "verbose": 0,
      "warm_start": false,
      "class_weight": null,
      "ccp_alpha": 0.0,
      "max_samples": null,
      "monotonic_cst": null
    }
  },

Cross Validation

Cross validation can be enabled to be applied to balanced training data only. This gives an estimate of how well a classifier can learn from the balanced data and generalise across different parts of that balanced dataset.

To apply cross validation with balanced datasets, apply a cross validation number with:

balancr configure-evaluation 5

This will perform cross validation with 5 folds, meaning each balanced dataset will be split into 5 folds, and will train the selected classifiers in 5 rounds.

An average of these round's results will be retrieved

Learning Curves

The same process mentioned in cross validation above is applied to generating learning curves.

Learning curves help us visualise each model's performance when being trained on increasing amounts of data.

These learning curves are generated using the balanced datasets chosen by the user

To configure learning curves:

balancr configure-evaluations --learning-curve-folds 5 --learning-curve-points 10

This will set the number of cross validation folds and number of points to plot on the learning curves

Requirements

• Python >= 3.8
• NumPy >= 1.21.0
• pandas >= 1.3.0
• scikit-learn >= 1.0.0
• matplotlib >= 3.4.0
• seaborn >= 0.11.0
• imbalanced-learn >= 0.8.0
• openpyxl >= 3.0.0
• colorama >= 0.4.4

Future Plans

• More visualisation options
• Collecting balancer and classifier times, other than only displaying in logs
• Saving results as runs go along, rather than retrieving all results at end of run

Author

Conor Doherty, [e-mail]