Skip to content

diver.md

Latest commit

 

History

History
144 lines (107 loc) · 3.58 KB

diver.md

File metadata and controls

144 lines (107 loc) · 3.58 KB

Diver

Diver implements a rule-based explainer that discovers discriminative patterns through combinatorial optimization. Primarily used for anomaly detection or class differentiation in classification tasks.

Class Definition

class mindxlib.explainer.Diver(
    label_col: str,
    label_val: Any,
    pos_beta: float = 1.5,
    overlap_beta_: float = 0.2,
    complexity_cost: float = 0.00001,
    dim_list: Optional[List[str]] = None,
    sup_ratio: float = 0.01,
    write_model: bool = False,
    disable_log: bool = True,
    cache_ind: bool = False
)

Parameters

  • label_col : str

    • Column name for the label in the dataset

  • label_val : Any

    • Target class value (e.g., 1 for anomalies)

  • pos_beta : float, default=1.5

    • Weight coefficient for positive samples

  • overlap_beta_ : float, default=0.2

    • Penalty coefficient for rule overlap

  • complexity_cost : float, default=0.00001

    • Penalty coefficient for rule length

  • dim_list : List[str], optional, default=None

    • Feature columns for rule mining
    • Uses all non-label columns if None

  • sup_ratio : float, default=0.01

    • Minimum support ratio (relative to target class count)

  • write_model : bool, default=False

    • Whether to save optimization models (for debugging)

  • disable_log : bool, default=True

    • Disable logging output

  • cache_ind : bool, default=False

    • Enable intermediate result caching

Methods

fit()

Learn patterns from training data.

def fit(
    X: Union[pd.DataFrame, np.ndarray],
    y: Union[pd.Series, np.ndarray],
    X_columns: Optional[List[str]] = None,
    y_column: Optional[str] = None,
    default_label: Optional[Any] = None
) -> Diver

Parameters:

  • X : pd.DataFrame or np.ndarray

    • Feature matrix of shape (n_samples, n_features)

  • y : pd.Series or np.ndarray

    • Label vector of shape (n_samples,)

  • X_columns : List[str], optional

    • Feature column names (required when X is ndarray)

  • y_column : str, optional

    • Label column name (required when y is ndarray)

  • default_label : Any, optional

    • Manually specified default rule label

Returns:

  • Fitted Diver instance

predict()

Make predictions on new data.

def predict(X_test: Union[pd.DataFrame, np.ndarray]) -> pd.Series

Parameters:

  • X_test : pd.DataFrame or np.ndarray
    • Test data of shape (n_samples, n_features)

Returns:

  • pd.Series of predicted labels, shape (n_samples,)

Examples

Basic Usage

from mindxlib import Diver
import pandas as pd
import numpy as np

# Create binary classification example data (anomaly detection)
X = pd.DataFrame({
    'age': ['young', 'middle-aged', 'old', 'middle-aged', 'old'],
    'income': ['high', 'medium', 'low', 'low', 'medium'],
})
y = pd.Series([0, 1, 1, 0, 1], name='is_anomaly')  # 1 indicates anomaly

# Initialize Diver
explainer = Diver(
    label_col='is_anomaly',
    label_val=1,  # We're interested in rules for anomalies
    pos_beta=1.5,  # Higher weight on catching anomalies
    overlap_beta_=0.3,  # Moderate penalty for rule overlap
    complexity_cost=0.001,  # Small penalty for longer rules
    dim_list=['age', 'income'],  # Focus on these features
    sup_ratio=0.2  # Minimum 20% support in anomaly class
)

# Fit the Model
explainer.fit(X, y)

# Make predictions on new data
test_data = pd.DataFrame({
    'age': ['young', 'middle-aged', 'old'],
    'income': ['high', 'medium', 'low']
})
fraud_predictions = explainer.predict(test_data)

# Show the rules in pure text
explainer.rules.show()