backend/utils/metrics.py

import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import tempfile
from typing import Tuple, Optional
from sklearn.metrics import confusion_matrix, accuracy_score, precision_score, recall_score, f1_score


def create_confusion_matrix_plot(
    cm: np.ndarray, 
    accuracy: float, 
    labels: list = ['No', 'Yes']
) -> str:
    """
    Create a confusion matrix plot and save it to a temporary file.
    
    Args:
        cm: Confusion matrix array
        accuracy: Accuracy score
        labels: Labels for the confusion matrix
        
    Returns:
        Path to the saved plot file
    """
    plt.figure(figsize=(6, 5))
    sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', xticklabels=labels, yticklabels=labels)
    plt.title(f'Confusion Matrix (Accuracy: {accuracy:.1%})')
    plt.ylabel('Ground Truth')
    plt.xlabel('Model Prediction')
    
    temp_file = tempfile.mktemp(suffix='.png')
    plt.savefig(temp_file, dpi=150, bbox_inches='tight')
    plt.close()
    
    return temp_file


def create_accuracy_table(df: pd.DataFrame) -> Tuple[pd.DataFrame, str, pd.DataFrame]:
    """
    Create accuracy metrics table and confusion matrix from results dataframe.
    
    Args:
        df: DataFrame with 'Ground Truth' and 'Binary Output' columns
        
    Returns:
        Tuple of (metrics_df, confusion_matrix_plot_path, confusion_matrix_values_df)
        
    Raises:
        ValueError: If insufficient data for binary classification
    """
    df_copy = df.copy()
    
    # Get unique values from both Ground Truth and Binary Output
    # Convert to string first, then apply .str operations
    ground_truth_values = df_copy['Ground Truth'].dropna().astype(str).str.lower().unique()
    binary_output_values = df_copy['Binary Output'].dropna().astype(str).str.lower().unique()
    
    # Combine and get all unique values
    all_values = set(list(ground_truth_values) + list(binary_output_values))
    all_values = [v for v in all_values if v.strip()]  # Remove empty strings
    
    if len(all_values) < 2:
        raise ValueError("Need at least 2 different values for binary classification")
    
    # Sort values to ensure consistent mapping (alphabetical order)
    sorted_values = sorted(all_values)
    
    # Create mapping: first value (alphabetically) = 0, second = 1
    # This ensures consistent mapping regardless of order in data
    value_mapping = {sorted_values[0]: 0}
    if len(sorted_values) >= 2:
        value_mapping[sorted_values[1]] = 1
    
    # If there are more than 2 values, map the rest to 1 (positive class)
    for i in range(2, len(sorted_values)):
        value_mapping[sorted_values[i]] = 1
    
    print(f"Detected binary mapping: {value_mapping}")
    
    # Apply mapping - convert to string first, then apply .str operations
    df_copy['Ground Truth Binary'] = df_copy['Ground Truth'].astype(str).str.lower().map(value_mapping)
    df_copy['Binary Output Binary'] = df_copy['Binary Output'].astype(str).str.lower().map(value_mapping)
    
    # Remove rows where either ground truth or binary output is NaN
    df_copy = df_copy.dropna(subset=['Ground Truth Binary', 'Binary Output Binary'])
    
    if len(df_copy) == 0:
        raise ValueError("No valid data for accuracy calculation after mapping. Check that Ground Truth and Binary Output contain valid binary values.")
    
    # Calculate metrics
    cm = confusion_matrix(df_copy['Ground Truth Binary'], df_copy['Binary Output Binary'])
    accuracy = accuracy_score(df_copy['Ground Truth Binary'], df_copy['Binary Output Binary'])
    precision = precision_score(df_copy['Ground Truth Binary'], df_copy['Binary Output Binary'], zero_division=0)
    recall = recall_score(df_copy['Ground Truth Binary'], df_copy['Binary Output Binary'], zero_division=0)
    f1 = f1_score(df_copy['Ground Truth Binary'], df_copy['Binary Output Binary'], zero_division=0)
    
    # Create metrics dataframe
    metrics_data = [
        ["Accuracy", f"{accuracy:.3f}"],
        ["Precision", f"{precision:.3f}"],
        ["Recall", f"{recall:.3f}"],
        ["F1 Score", f"{f1:.3f}"],
        ["Total Samples", f"{len(df_copy)}"]
    ]
    metrics_df = pd.DataFrame(metrics_data, columns=["Metric", "Value"])
    
    # Create labels for confusion matrix based on detected values
    # Find the original case versions of the labels
    original_labels = []
    for mapped_val in sorted([k for k, v in value_mapping.items() if v in [0, 1]]):
        # Find original case version from the data
        original_case = None
        for val in df_copy['Ground Truth'].dropna():
            if str(val).lower() == mapped_val:
                original_case = str(val)
                break
        if original_case is None:
            for val in df_copy['Binary Output'].dropna():
                if str(val).lower() == mapped_val:
                    original_case = str(val)
                    break
        original_labels.append(original_case if original_case else mapped_val.title())
    
    # Ensure we have exactly 2 labels
    if len(original_labels) < 2:
        original_labels = ['Class 0', 'Class 1']
    
    cm_plot_path = create_confusion_matrix_plot(cm, accuracy, original_labels)
    
    # Confusion matrix values table
    if cm.shape == (2, 2):
        tn, fp, fn, tp = cm.ravel()
        cm_values = pd.DataFrame(
            [[tn, fp], [fn, tp]],
            columns=[f"Predicted {original_labels[0]}", f"Predicted {original_labels[1]}"],
            index=[f"Actual {original_labels[0]}", f"Actual {original_labels[1]}"]
        )
    else:
        cm_values = pd.DataFrame(cm)
    
    return metrics_df, cm_plot_path, cm_values


def save_dataframe_to_csv(df: pd.DataFrame) -> Optional[str]:
    """
    Save dataframe to a temporary CSV file.
    
    Args:
        df: DataFrame to save
        
    Returns:
        Path to saved CSV file or None if failed
    """
    if df is None or df.empty:
        return None
        
    temp_file = tempfile.mktemp(suffix='.csv')
    df.to_csv(temp_file, index=False)
    return temp_file