Overview This guide walks through using Vishwas to detect and address bias in machine learning models, with a focus on Indian regulatory requirements and protected attributes.
Prerequisites
Rotavision account with Vishwas access
Model predictions and ground truth labels
Identified protected attributes (gender, region, etc.)
Step 1: Prepare Your Data import pandas as pd from rotavision import Rotavision client = Rotavision() # Load your dataset df = pd.read_csv( "loan_applications.csv" ) # Prepare for analysis dataset = { "features" : df.columns.tolist(), "data" : df.values.tolist(), "predictions" : df[ "model_prediction" ].tolist(), "actuals" : df[ "approved" ].tolist(), "protected_attributes" : [ "gender" , "region" , "age_group" ] }
Step 2: Run Fairness Analysis analysis = client.vishwas.analyze( model_id = "loan-approval-v2" , dataset = dataset, metrics = [ "demographic_parity" , "equalized_odds" , "equal_opportunity" , "calibration" ], thresholds = { "demographic_parity" : 0.80 , # 80% threshold "equalized_odds" : 0.80 } ) print ( f "Overall Fairness Score: { analysis.overall_score :.2f} " ) print ( f "Bias Detected: { analysis.bias_detected } " )
Step 3: Interpret Results for metric in analysis.metrics: status_icon = "✅" if metric.status == "pass" else "⚠️" print ( f " { status_icon } { metric.name } : { metric.value :.3f} (threshold: { metric.threshold } )" ) if metric.affected_groups: print ( f " Affected groups: { ', ' .join(metric.affected_groups) } " )
Example Output ⚠️ demographic_parity: 0.72 (threshold: 0.80) Affected groups: region:rural, gender:female ✅ equalized_odds: 0.85 (threshold: 0.80) ✅ equal_opportunity: 0.88 (threshold: 0.80) ✅ calibration: 0.91 (threshold: 0.80)
Step 4: Review Recommendations for rec in analysis.recommendations: print ( f "[ { rec.severity.upper() } ] { rec.message } " ) print ( f " Action: { rec.action } " ) print ( f " Expected impact: { rec.impact_estimate } " ) print ()
Step 5: Generate Compliance Report report = client.vishwas.generate_report( analysis_id = analysis.id, template = "rbi" , # RBI compliance format format = "pdf" , metadata = { "model_name" : "Loan Approval Model v2" , "model_owner" : "Credit Risk Team" , "review_date" : "2026-02-01" } ) # Download report print ( f "Report URL: { report.download_url } " )
Common Mitigation Strategies
Balance training data across protected groups: from sklearn.utils import resample # Oversample minority groups df_majority = df[df.gender == 'male' ] df_minority = df[df.gender == 'female' ] df_minority_upsampled = resample( df_minority, replace = True , n_samples = len (df_majority) )
Apply different decision thresholds per group: # Post-processing threshold adjustment thresholds = { "urban" : 0.50 , "rural" : 0.45 , # Lower threshold for disadvantaged group }
Remove or transform proxy features: # Remove features highly correlated with protected attributes corr = df[[ 'prediction' , 'pincode' , 'region' ]].corr() # Consider removing 'pincode' if it's a proxy for region
Indian Regulatory Context RBI Guidelines The RBI requires fairness analysis for AI/ML models used in:
Credit scoring and lending decisions
Customer segmentation
Fraud detection
Key requirements:
Document protected attributes considered
Quantify disparate impact
Implement ongoing monitoring
Protected Attributes in India Common protected attributes to analyze:
Gender : Male, Female, OtherRegion : Urban, Semi-urban, RuralState : Geographic bias across statesLanguage : Language preference as proxyAge : Age-based discrimination
Caste and religion are highly sensitive in India. Consult legal counsel before including in analysis, even for bias detection purposes.
Continuous Monitoring Set up Guardian to monitor fairness drift in production:
monitor = client.guardian.create_monitor( model_id = "loan-approval-v2" , name = "Fairness Monitor" , metrics = [ "prediction_drift" ], alerts = [ { "metric" : "prediction_drift" , "threshold" : 0.15 , "severity" : "warning" , "group_by" : "region" # Monitor drift per region } ] )
Next Steps 
