predict.py

"""
Insurance Claims Decision Prediction Script
Provides predictions with required output format
"""

import joblib
import json
import numpy as np
import hashlib
from datetime import datetime

class ClaimsDecisionPredictor:
    def __init__(self, model_path='model_artifacts'):
        """Load model and metadata"""
        self.model = joblib.load(f'{model_path}/claims_decision_model.pkl')
        self.accident_encoder = joblib.load(f'{model_path}/accident_type_encoder.pkl')
        self.police_encoder = joblib.load(f'{model_path}/police_report_encoder.pkl')
        
        with open(f'{model_path}/metadata.json', 'r') as f:
            self.metadata = json.load(f)
    
    def predict(self, claim_data):
        """
        Make prediction with required output format
        
        Args:
            claim_data (dict): Dictionary with keys:
                - claim_amount (float)
                - vehicle_age (int)
                - accident_type (str)
                - police_report (str: 'yes' or 'no')
                - repair_estimate (float)
                - prior_claims (int)
        
        Returns:
            dict: {
                "decision": "<approve|review|reject>",
                "confidence": <0.0-1.0>,
                "top_factors": ["factor_1", "factor_2", "factor_3"]
            }
        """
        # Encode categorical variables
        accident_encoded = self.accident_encoder.transform([claim_data['accident_type']])[0]
        police_encoded = self.police_encoder.transform([claim_data['police_report']])[0]
        
        # Create derived features
        claim_to_estimate_ratio = claim_data['claim_amount'] / (claim_data['repair_estimate'] + 1)
        high_claim_flag = int(claim_data['claim_amount'] > 15000)
        old_vehicle_flag = int(claim_data['vehicle_age'] > 10)
        multiple_claims_flag = int(claim_data['prior_claims'] > 2)
        
        # Prepare feature vector
        features = np.array([[
            claim_data['claim_amount'],
            claim_data['vehicle_age'],
            accident_encoded,
            police_encoded,
            claim_data['repair_estimate'],
            claim_data['prior_claims'],
            claim_to_estimate_ratio,
            high_claim_flag,
            old_vehicle_flag,
            multiple_claims_flag
        ]])
        
        # Get prediction and probabilities
        decision = self.model.predict(features)[0]
        probabilities = self.model.predict_proba(features)[0]
        
        # Get confidence (probability of predicted class)
        decision_idx = list(self.model.classes_).index(decision)
        confidence = float(probabilities[decision_idx])
        
        # Get top factors based on feature importance for this decision
        top_factors = self._get_top_factors(features[0], decision)
        
        return {
            "decision": decision,
            "confidence": round(confidence, 4),
            "top_factors": top_factors
        }
    
    def _get_top_factors(self, feature_values, decision):
        """Extract top 3 factors influencing the decision"""
        decision_idx = list(self.model.classes_).index(decision)
        coefficients = self.model.coef_[decision_idx]
        
        # Calculate contribution of each feature (coefficient * feature_value)
        contributions = np.abs(coefficients * feature_values)
        
        # Get top 3 feature indices
        top_indices = np.argsort(contributions)[-3:][::-1]
        
        # Map to readable feature names
        readable_names = self.metadata['feature_names_readable']
        top_factors = [readable_names[idx] for idx in top_indices]
        
        return top_factors
    
    def predict_with_logging(self, claim_data, log_file='inference_logs.jsonl'):
        """Make prediction and log the inference"""
        prediction = self.predict(claim_data)
        
        # Create log entry
        log_entry = {
            "timestamp": datetime.now().isoformat(),
            "model_version": self.metadata['version'],
            "input_hash": hashlib.md5(json.dumps(claim_data, sort_keys=True).encode()).hexdigest(),
            "decision": prediction['decision'],
            "confidence": prediction['confidence']
        }
        
        # Append to log file
        with open(log_file, 'a') as f:
            f.write(json.dumps(log_entry) + '\n')
        
        return prediction


# Test the predictor
if __name__ == "__main__":
    predictor = ClaimsDecisionPredictor()
    
    # Test case 1: Should approve
    test_claim_1 = {
        "claim_amount": 3500.0,
        "vehicle_age": 3,
        "accident_type": "collision",
        "police_report": "yes",
        "repair_estimate": 3200.0,
        "prior_claims": 0
    }
    
    # Test case 2: Should review
    test_claim_2 = {
        "claim_amount": 12000.0,
        "vehicle_age": 8,
        "accident_type": "theft",
        "police_report": "no",
        "repair_estimate": 8000.0,
        "prior_claims": 2
    }
    
    # Test case 3: Should reject
    test_claim_3 = {
        "claim_amount": 25000.0,
        "vehicle_age": 15,
        "accident_type": "vandalism",
        "police_report": "no",
        "repair_estimate": 5000.0,
        "prior_claims": 5
    }
    
    print("="*60)
    print("TESTING CLAIMS DECISION PREDICTOR")
    print("="*60)
    
    for i, test_claim in enumerate([test_claim_1, test_claim_2, test_claim_3], 1):
        print(f"\nTest Case {i}:")
        print(f"Input: {json.dumps(test_claim, indent=2)}")
        
        result = predictor.predict_with_logging(test_claim)
        print(f"\nPrediction:")
        print(json.dumps(result, indent=2))
        print("-"*60)
    
    print("\n✅ Prediction tests complete!")
    print("Logs saved to: inference_logs.jsonl")