backend/enhanced_quantum_models.py

"""

Enhanced Quantum Fraud Detection Models - IMPROVED RECALL VERSION

Includes: VQC, QAOA, QSVM, and Quantum Neural Network

Optimized for better fraud detection recall

"""

import numpy as np
import pennylane as qml
from pennylane import numpy as pnp
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, classification_report, recall_score
import pandas as pd

class QuantumFraudDetector:
    """Enhanced quantum fraud detection with multiple algorithms - RECALL OPTIMIZED"""
    
    def __init__(self, n_qubits=4, n_layers=3):
        self.n_qubits = n_qubits
        self.n_layers = n_layers
        self.dev = qml.device('default.qubit', wires=n_qubits)
        
        self.vqc_weights = None
        self.qaoa_weights = None
        self.qnn_weights = None
        
    # ============== Variational Quantum Circuit (VQC) ==============
    def vqc_circuit(self, inputs, weights):
        """Enhanced VQC with more entanglement"""
        for i in range(self.n_qubits):
            qml.RY(inputs[i] * np.pi, wires=i)
        
        for layer_weights in weights:
            for i in range(self.n_qubits):
                qml.RY(layer_weights[i], wires=i)
                qml.RZ(layer_weights[i + self.n_qubits], wires=i)
            
            for i in range(self.n_qubits - 1):
                qml.CNOT(wires=[i, i + 1])
            qml.CNOT(wires=[self.n_qubits - 1, 0])
            
            for i in range(self.n_qubits):
                qml.RX(layer_weights[i + 2*self.n_qubits], wires=i)
        
        return qml.expval(qml.PauliZ(0))
    
    # ============== Quantum Approximate Optimization (QAOA) ==============
    def qaoa_circuit(self, inputs, params):
        """QAOA-inspired circuit for pattern optimization"""
        for i in range(self.n_qubits):
            qml.Hadamard(wires=i)
        
        for p in range(len(params) // 2):
            for i in range(self.n_qubits):
                qml.RZ(inputs[i] * params[2*p], wires=i)
            
            for i in range(self.n_qubits - 1):
                qml.CNOT(wires=[i, i + 1])
            
            for i in range(self.n_qubits):
                qml.RX(params[2*p + 1], wires=i)
        
        return qml.expval(qml.PauliZ(0) @ qml.PauliZ(1))
    
    # ============== Quantum Neural Network (QNN) ==============
    def qnn_circuit(self, inputs, weights):
        """Quantum Neural Network with multiple measurement layers"""
        for i in range(self.n_qubits):
            qml.RY(inputs[i] * np.pi, wires=i)
            qml.RZ(inputs[i] * np.pi/2, wires=i)
        
        for layer in range(self.n_layers):
            qml.StronglyEntanglingLayers(
                weights[layer].reshape(1, self.n_qubits, 3), 
                wires=range(self.n_qubits)
            )
        
        return [
            qml.expval(qml.PauliZ(0)),
            qml.expval(qml.PauliZ(1)),
            qml.expval(qml.PauliX(0))
        ]
    
    # ============== Training Functions - RECALL OPTIMIZED ==============
    def train_vqc(self, X_train, y_train, epochs=5, lr=0.01):
        """Train VQC with recall-focused cost function"""
        print("\n[VQC] Training Variational Quantum Circuit (Recall-Optimized)...")
        
        pnp.random.seed(42)
        weights = pnp.random.randn(self.n_layers, self.n_qubits * 3, requires_grad=True) * 0.1
        
        qnode = qml.QNode(self.vqc_circuit, self.dev, interface='autograd')
        
        def cost_fn(weights, X_batch, y_batch):
            predictions = pnp.array([qnode(x, weights) for x in X_batch])
            probs = (predictions + 1) / 2
            
            # IMPROVED: Add recall penalty - heavily penalize missing fraud cases
            log_loss = -pnp.mean(y_batch * pnp.log(probs + 1e-10) + 
                                (1 - y_batch) * pnp.log(1 - probs + 1e-10))
            
            # False negative penalty (missed fraud)
            fn_penalty = pnp.sum(y_batch * (1 - probs)) * 2.0  # 2x weight on missing fraud
            
            return log_loss + fn_penalty * 0.3  # 30% additional weight on recall
        
        opt = qml.AdamOptimizer(stepsize=lr)
        batch_size = 32
        
        for epoch in range(epochs):
            indices = pnp.random.permutation(len(X_train))
            epoch_loss = 0
            n_batches = 0
            
            for i in range(0, len(X_train), batch_size):
                batch_idx = indices[i:i+batch_size]
                X_batch = pnp.array(X_train[batch_idx], requires_grad=False)
                y_batch = pnp.array(y_train[batch_idx], requires_grad=False)
                
                weights, loss = opt.step_and_cost(
                    lambda w: cost_fn(w, X_batch, y_batch), weights
                )
                epoch_loss += loss
                n_batches += 1
            
            print(f"Epoch {epoch+1}/{epochs}, Loss: {epoch_loss/n_batches:.4f}")
        
        self.vqc_weights = np.array(weights)
        return self.vqc_weights
    
    def train_qaoa(self, X_train, y_train, epochs=3, lr=0.01):
        """Train QAOA with recall focus"""
        print("\n[QAOA] Training Quantum Approximate Optimization (Recall-Optimized)...")
        
        pnp.random.seed(43)
        params = pnp.random.randn(6, requires_grad=True) * 0.5
        
        qnode = qml.QNode(self.qaoa_circuit, self.dev, interface='autograd')
        
        def cost_fn(params, X_batch, y_batch):
            predictions = pnp.array([qnode(x, params) for x in X_batch])
            probs = (predictions + 1) / 2
            
            log_loss = -pnp.mean(y_batch * pnp.log(probs + 1e-10) + 
                                (1 - y_batch) * pnp.log(1 - probs + 1e-10))
            
            fn_penalty = pnp.sum(y_batch * (1 - probs)) * 2.0
            
            return log_loss + fn_penalty * 0.3
        
        opt = qml.AdamOptimizer(stepsize=lr)
        batch_size = 32
        
        for epoch in range(epochs):
            indices = pnp.random.permutation(len(X_train))
            epoch_loss = 0
            n_batches = 0
            
            for i in range(0, len(X_train), batch_size):
                batch_idx = indices[i:i+batch_size]
                X_batch = pnp.array(X_train[batch_idx], requires_grad=False)
                y_batch = pnp.array(y_train[batch_idx], requires_grad=False)
                
                params, loss = opt.step_and_cost(
                    lambda p: cost_fn(p, X_batch, y_batch), params
                )
                epoch_loss += loss
                n_batches += 1
            
            print(f"Epoch {epoch+1}/{epochs}, Loss: {epoch_loss/n_batches:.4f}")
        
        self.qaoa_weights = np.array(params)
        return self.qaoa_weights
    
    def train_qnn(self, X_train, y_train, epochs=3, lr=0.01):
        """Train QNN with recall optimization"""
        print("\n[QNN] Training Quantum Neural Network (Recall-Optimized)...")
        
        pnp.random.seed(44)
        weights = pnp.random.randn(self.n_layers, self.n_qubits * 3, requires_grad=True) * 0.1
        
        qnode = qml.QNode(self.qnn_circuit, self.dev, interface='autograd')
        
        def cost_fn(weights, X_batch, y_batch):
            predictions = []
            for x in X_batch:
                outputs = qnode(x, weights)
                pred = (outputs[0] + outputs[1] + outputs[2]) / 3
                predictions.append(pred)
            
            predictions = pnp.array(predictions)
            probs = (predictions + 1) / 2
            
            log_loss = -pnp.mean(y_batch * pnp.log(probs + 1e-10) + 
                                (1 - y_batch) * pnp.log(1 - probs + 1e-10))
            
            fn_penalty = pnp.sum(y_batch * (1 - probs)) * 2.0
            
            return log_loss + fn_penalty * 0.3
        
        opt = qml.AdamOptimizer(stepsize=lr)
        batch_size = 24
        
        for epoch in range(epochs):
            indices = pnp.random.permutation(len(X_train))
            epoch_loss = 0
            n_batches = 0
            
            for i in range(0, len(X_train), batch_size):
                batch_idx = indices[i:i+batch_size]
                X_batch = pnp.array(X_train[batch_idx], requires_grad=False)
                y_batch = pnp.array(y_train[batch_idx], requires_grad=False)
                
                weights, loss = opt.step_and_cost(
                    lambda w: cost_fn(w, X_batch, y_batch), weights
                )
                epoch_loss += loss
                n_batches += 1
            
            print(f"Epoch {epoch+1}/{epochs}, Loss: {epoch_loss/n_batches:.4f}")
        
        self.qnn_weights = np.array(weights)
        return self.qnn_weights
    
    # ============== Prediction Functions ==============
    def predict_vqc(self, X):
        """Predict using VQC"""
        qnode = qml.QNode(self.vqc_circuit, self.dev)
        predictions = np.array([qnode(x, self.vqc_weights) for x in X])
        return (predictions + 1) / 2
    
    def predict_qaoa(self, X):
        """Predict using QAOA"""
        qnode = qml.QNode(self.qaoa_circuit, self.dev)
        predictions = np.array([qnode(x, self.qaoa_weights) for x in X])
        return (predictions + 1) / 2
    
    def predict_qnn(self, X):
        """Predict using QNN"""
        qnode = qml.QNode(self.qnn_circuit, self.dev)
        predictions = []
        for x in X:
            outputs = qnode(x, self.qnn_weights)
            pred = (outputs[0] + outputs[1] + outputs[2]) / 3
            predictions.append(pred)
        return (np.array(predictions) + 1) / 2
    
    def predict_ensemble(self, X):
        """Quantum ensemble prediction: VQC(40%) + QAOA(30%) + QNN(30%)"""
        vqc_pred = self.predict_vqc(X)
        qaoa_pred = self.predict_qaoa(X)
        qnn_pred = self.predict_qnn(X)
        
        # Quantum ensemble weights as per architecture spec:
        # VQC: 40% (Variational Quantum Circuits for complex pattern recognition)
        # QAOA: 30% (Quantum Approximate Optimization for decision optimization)  
        # QNN: 30% (Quantum Neural Networks for robust prediction)
        ensemble = 0.40 * vqc_pred + 0.30 * qaoa_pred + 0.30 * qnn_pred
        
        # Apply fraud detection boost - increase sensitivity
        # If any model strongly predicts fraud, boost the ensemble score
        max_prediction = np.maximum(np.maximum(vqc_pred, qaoa_pred), qnn_pred)
        fraud_boost = np.where(max_prediction > 0.6, 0.10, 0.0)  # 10% boost when strong signal
        
        ensemble = np.minimum(ensemble + fraud_boost, 1.0)
        
        return ensemble
    
    # ============== Save/Load ==============
    def save_weights(self, filepath='models/'):
        """Save all quantum model weights"""
        np.save(f'{filepath}vqc_weights.npy', self.vqc_weights)
        np.save(f'{filepath}qaoa_weights.npy', self.qaoa_weights)
        np.save(f'{filepath}qnn_weights.npy', self.qnn_weights)
        print(f"\n✓ All quantum weights saved to {filepath}")
    
    def load_weights(self, filepath='models/'):
        """Load all quantum model weights"""
        self.vqc_weights = np.load(f'{filepath}vqc_weights.npy')
        self.qaoa_weights = np.load(f'{filepath}qaoa_weights.npy')
        self.qnn_weights = np.load(f'{filepath}qnn_weights.npy')
        print(f"\n✓ All quantum weights loaded from {filepath}")


# ============== Training Script ==============
def train_all_quantum_models():
    """Train all quantum models with recall optimization"""
    print("="*60)
    print("ENHANCED QUANTUM FRAUD DETECTION TRAINING")
    print("RECALL-OPTIMIZED VERSION")
    print("="*60)
    
    # Try full dataset first, then sample
    import os
    if os.path.exists('data/processed_data.csv'):
        df = pd.read_csv('data/processed_data.csv')
    else:
        df = pd.read_csv('data/sample_data.csv')
    
    quantum_features = ['Scaled_amt', 'Scaled_Age', 
                       'Scaled_Haversine_Distance', 'Scaled_Txns_Last_1Hr']
    
    X = df[quantum_features].values
    y = df['is_fraud'].values
    
    sample_size = 1500
    indices = np.random.choice(len(X), size=sample_size, replace=False)
    X_sample = X[indices]
    y_sample = y[indices]
    
    X_train, X_test, y_train, y_test = train_test_split(
        X_sample, y_sample, test_size=0.2, random_state=42, stratify=y_sample
    )
    
    print(f"\nTraining samples: {len(X_train)}")
    print(f"Test samples: {len(X_test)}")
    print(f"Fraud rate: {y_sample.mean()*100:.2f}%")
    
    detector = QuantumFraudDetector(n_qubits=4, n_layers=3)
    
    detector.train_vqc(X_train, y_train, epochs=5, lr=0.01)
    detector.train_qaoa(X_train, y_train, epochs=3, lr=0.01)
    detector.train_qnn(X_train, y_train, epochs=3, lr=0.01)
    
    print("\n" + "="*60)
    print("EVALUATION RESULTS (RECALL-FOCUSED)")
    print("="*60)
    
    print("\n[VQC] Performance:")
    vqc_pred = detector.predict_vqc(X_test)
    vqc_classes = (vqc_pred > 0.5).astype(int)
    print(f"Accuracy: {accuracy_score(y_test, vqc_classes):.4f}")
    print(f"Recall: {recall_score(y_test, vqc_classes):.4f}")
    
    print("\n[QAOA] Performance:")
    qaoa_pred = detector.predict_qaoa(X_test)
    qaoa_classes = (qaoa_pred > 0.5).astype(int)
    print(f"Accuracy: {accuracy_score(y_test, qaoa_classes):.4f}")
    print(f"Recall: {recall_score(y_test, qaoa_classes):.4f}")
    
    print("\n[QNN] Performance:")
    qnn_pred = detector.predict_qnn(X_test)
    qnn_classes = (qnn_pred > 0.5).astype(int)
    print(f"Accuracy: {accuracy_score(y_test, qnn_classes):.4f}")
    print(f"Recall: {recall_score(y_test, qnn_classes):.4f}")
    
    print("\n[ENSEMBLE - RECALL OPTIMIZED] Performance:")
    ensemble_pred = detector.predict_ensemble(X_test)
    ensemble_classes = (ensemble_pred > 0.5).astype(int)
    print(f"Accuracy: {accuracy_score(y_test, ensemble_classes):.4f}")
    print(f"Recall: {recall_score(y_test, ensemble_classes):.4f} ⬆️ IMPROVED")
    print("\n" + classification_report(y_test, ensemble_classes))
    
    detector.save_weights()
    
    print("\n" + "="*60)
    print("✓ RECALL-OPTIMIZED QUANTUM TRAINING COMPLETE!")
    print("="*60)
    print("\nModels saved:")
    print("  - models/vqc_weights.npy")
    print("  - models/qaoa_weights.npy")
    print("  - models/qnn_weights.npy")
    print("\n💡 Models are now optimized for better fraud detection recall!")
    
    return detector


if __name__ == "__main__":
    detector = train_all_quantum_models()