Select Page

Statistical Basis for Bias Identification and Correction Code

import os
import pandas as pd
import numpy as np
from scipy.stats import chisquare, skew, kurtosis, norm
from statsmodels.stats.proportion import proportion_confint
from statsmodels.stats.contingency_tables import Table2x2

# Load your data
objects = pd.read_csv('clarifai_long_format.csv')  # must include columns: image, clarifai_category, object_label

# Get category sizes
cat_sizes = objects.groupby('clarifai_category')['image'].nunique().sort_values()
cat_counts = cat_sizes.values
categories = cat_sizes.index.tolist()
total_images = cat_counts.sum()
num_categories = len(cat_counts)

print(f"Total images: {total_images}\nNumber of categories: {num_categories}")

# Category stats
print(f"Category size range: {cat_counts.min()}–{cat_counts.max()} images")
imbalance_ratio = cat_counts.max() / cat_counts.min()
print(f"Imbalance ratio: {imbalance_ratio:.2f}:1")
mean = cat_counts.mean()
std = cat_counts.std()
ci95 = norm.interval(0.95, loc=mean, scale=std/np.sqrt(num_categories))
print(f"Mean size: {mean:.1f} ± {std:.1f} (95% CI: {ci95[0]:.1f}–{ci95[1]:.1f})")

# Gini coefficient
def gini(x):
    diffsum = np.sum(np.abs(np.subtract.outer(x, x)))
    return diffsum / (2 * len(x)**2 * np.mean(x))
gini_coef = gini(cat_counts)
print(f"Gini coefficient: {gini_coef:.3f}")

# Chi-square for category evenness
expected = np.ones(num_categories) * (total_images / num_categories)
chi2, p = chisquare(cat_counts, expected)
print(f"Chi-square goodness of fit: χ² = {chi2:.3f}, p = {p:.3g}")

# Cohen's w
cohen_w = np.sqrt(chi2 / total_images)
print(f"Cohen’s w: {cohen_w:.3f}")

# Variance, CV, skew, kurtosis
var = cat_counts.var()
cv = 100 * std / mean
sk = skew(cat_counts)
ku = kurtosis(cat_counts)
print(f"Variance in category sizes: σ² = {var:.1f}")
print(f"Coefficient of variation: {cv:.1f}%")
print(f"Skewness: {sk:.2f}")
print(f"Kurtosis: {ku:.2f}")

# Statistical power ratio and confidence interval width difference
stat_power_ratio = cat_counts.max() / cat_counts.min()
ci_diff = 100 * (1/np.sqrt(cat_counts.min()) - 1/np.sqrt(cat_counts.max())) / (1/np.sqrt(cat_counts.max()))
print(f"Statistical power ratio: {stat_power_ratio:.2f}:1")
print(f"Confidence interval precision: {ci_diff:.1f}% difference between smallest and largest categories")

# Object-level disparity: frequency of "people" in each category
people_counts = objects[objects['object_label'] == 'people'].groupby('clarifai_category')['object_label'].count().reindex(categories, fill_value=0)
expected_people = people_counts.sum() / num_categories

# Digital & natural categories (adjust for your naming)
digital_cats = ['technology_addiction', 'doomscrolling', 'social_media_addiction']
natural_cats = ['nature', 'instinct', 'artificial']

for cat in digital_cats + natural_cats:
    obs = people_counts.get(cat, 0)
    freq_ratio = obs / expected_people if expected_people > 0 else np.nan
    # Binomial proportion CI for observed/expected
    ci_lo, ci_hi = proportion_confint(obs, people_counts.sum(), alpha=0.05, method='wilson')
    print(f"{cat}: {freq_ratio:.2f}× expected frequency (95% CI: {ci_lo/expected_people:.2f}–{ci_hi/expected_people:.2f})")

# Chi-square for object distribution (people x 2 categories: digital vs non-digital)
objects['is_digital'] = objects['clarifai_category'].isin(digital_cats)
contingency = pd.crosstab(objects['is_digital'], objects['object_label'] == 'people')
chi2_obj, p_obj = chisquare(contingency.values.flatten())
print(f"Chi-square test for object distribution: χ² = {chi2_obj:.1f}, p < {p_obj:.3g}")

# Cramér’s V
def cramers_v(confusion_matrix):
    chi2 = chisquare(confusion_matrix.flatten())[0]
    n = confusion_matrix.sum()
    return np.sqrt(chi2 / (n * (min(confusion_matrix.shape)-1)))
cramers_v_val = cramers_v(contingency.values)
print(f"Cramér’s V: {cramers_v_val:.2f}")

# Distribution shape
print(f"Distribution: {'Right-skewed' if sk > 0 else 'Left-skewed' if sk < 0 else 'Symmetric'} with {'platykurtic' if ku < 0 else 'leptokurtic' if ku > 0 else 'mesokurtic'} distribution")

output_lines = []

def log(s):
    print(s)
    output_lines.append(str(s))

# -- Then, REPLACE every 'print(' line with 'log(' instead --

# Example:
log(f"Total images: {total_images}\nNumber of categories: {num_categories}")
log(f"Category size range: {cat_counts.min()}–{cat_counts.max()} images")
log(f"Imbalance ratio: {imbalance_ratio:.2f}:1")
log(f"Mean size: {mean:.1f} ± {std:.1f} (95% CI: {ci95[0]:.1f}–{ci95[1]:.1f})")
log(f"Gini coefficient: {gini_coef:.3f}")
log(f"Chi-square goodness of fit: χ² = {chi2:.3f}, p = {p:.3g}")
log(f"Cohen’s w: {cohen_w:.3f}")
log(f"Variance in category sizes: σ² = {var:.1f}")
log(f"Coefficient of variation: {cv:.1f}%")
log(f"Skewness: {sk:.2f}")
log(f"Kurtosis: {ku:.2f}")
log(f"Statistical power ratio: {stat_power_ratio:.2f}:1")
log(f"Confidence interval precision: {ci_diff:.1f}% difference between smallest and largest categories")

for cat in digital_cats + natural_cats:
    obs = people_counts.get(cat, 0)
    freq_ratio = obs / expected_people if expected_people > 0 else np.nan
    log(f"{cat}: {freq_ratio:.2f}× expected frequency")

# -- and for Cramér's V --
from scipy.stats import chi2_contingency
chi2_obj, p_obj, dof_obj, expected_obj = chi2_contingency(contingency)
n = contingency.values.sum()
k = min(contingency.shape)
cramers_v_val = np.sqrt(chi2_obj / (n * (k-1)))

log(f"Chi-square test for object distribution: χ² = {chi2_obj:.1f}, p = {p_obj:.3g}")
log(f"Cramér’s V: {cramers_v_val:.2f}")

log(f"Chi-square test for object distribution: χ² = {chi2_obj:.1f}, p < {p_obj:.3g}")
log(f"Cramér’s V: {cramers_v_val:.2f}")
log(f"Distribution: {'Right-skewed' if sk > 0 else 'Left-skewed' if sk < 0 else 'Symmetric'} with {'platykurtic' if ku < 0 else 'leptokurtic' if ku > 0 else 'mesokurtic'} distribution")

# --- Save to txt file at the end ---
with open('bias_stats_report.txt', 'w') as f:
    for line in output_lines:
        f.write(line + '\n')

print("Results saved to bias_stats_report.txt")


←→