Added more Models

app.py

@@ -2,58 +2,116 @@ import streamlit as st
 import pandas as pd
 import joblib
 import json
+import numpy as np
 from datetime import datetime
+from sklearn.neighbors import LocalOutlierFactor
 
-# Load trained model
-iso_forest = joblib.load("isolation_forest_model.pkl")
 
-# Load location mapping
-with open("location_mapping.json", "r") as f:
-    location_mapping = json.load(f)
+# ✅ Tabs for Application & Model Evaluation
+app, model_eval = st.tabs(["Application", "📊 Model Evaluation"])
 
-# Manual mapping for categorical variables
-transaction_type_mapping = {"Debit": 0, "Credit": 1}
-channel_mapping = {"ATM": 0, "Online": 1, "Branch": 2}
-day_of_week_mapping = {"Monday": 0, "Tuesday": 1, "Wednesday": 2, "Thursday": 3, "Friday": 4, "Saturday": 5, "Sunday": 6}
+# ---------------- APPLICATION TAB ---------------- #
+with app:
+    # Load trained models
+    iso_forest = joblib.load("isolation_forest_model.pkl")
+    one_class_svm = joblib.load("one_class_svm_model.pkl")
+    lof_model = joblib.load("local_outlier_factor_model.pkl")  # LOF model trained earlier
+    lof_threshold = joblib.load("lof_threshold.pkl")  # Precomputed threshold for LOF
 
-st.title("Anomaly Detection for Bank Transactions")
+    # Load location mapping
+    with open("location_mapping.json", "r") as f:
+        location_mapping = json.load(f)
 
+    # Manual mapping for categorical variables
+    transaction_type_mapping = {"Debit": 0, "Credit": 1}
+    channel_mapping = {"ATM": 0, "Online": 1, "Branch": 2}
+    day_of_week_mapping = {"Monday": 0, "Tuesday": 1, "Wednesday": 2, "Thursday": 3, "Friday": 4, "Saturday": 5, "Sunday": 6}
 
+    st.title("Anomaly Detection for Bank Transactions")
 
-# User inputs
-date = st.date_input("Select Transaction Date")
-time = st.time_input("Select Transaction Time")
-location = st.selectbox("Select Location", options=list(location_mapping.keys()))
-transaction_type = st.radio("Transaction Type", options=["Debit", "Credit"])
-channel = st.radio("Transaction Channel", options=["ATM", "Online", "Branch"])
-transaction_duration = st.slider("Transaction Duration (seconds)", min_value=0, max_value=600, value=30)
-login_attempts = st.number_input("Login Attempts", min_value=0)
-transaction_amount = st.number_input("Transaction Amount", min_value=0.0, format="%.2f")
+    # Sidebar for model selection
+    model_choice = st.sidebar.radio("Select Anomaly Detection Model", ["Isolation Forest", "One-Class SVM", "Local Outlier Factor"])
 
-if st.button("Check for Anomaly"):
-    # Convert date to day of the week
-    day_of_week = day_of_week_mapping[date.strftime('%A')]
+    # User inputs
+    date = st.date_input("Select Transaction Date")
+    time = st.time_input("Select Transaction Time")
+    location = st.selectbox("Select Location", options=list(location_mapping.keys()))
+    transaction_type = st.radio("Transaction Type", options=["Debit", "Credit"])
+    channel = st.radio("Transaction Channel", options=["ATM", "Online", "Branch"])
+    transaction_duration = st.slider("Transaction Duration (seconds)", min_value=0, max_value=600, value=30)
+    login_attempts = st.number_input("Login Attempts", min_value=0)
+    transaction_amount = st.number_input("Transaction Amount", min_value=0.0, format="%.2f")
+
+    if st.button("Check for Anomaly"):
+        # Convert date to day of the week
+        day_of_week = day_of_week_mapping[date.strftime('%A')]
+        
+        # Convert time to total seconds since midnight
+        total_seconds = time.hour * 3600 + time.minute * 60
+        
+        # Convert categorical values to numeric
+        location_encoded = location_mapping.get(location, -1)  # Default to -1 if not found
+        transaction_type_encoded = transaction_type_mapping[transaction_type]
+        channel_encoded = channel_mapping[channel]
+        
+        # Ensure the order of features matches training
+        input_data = pd.DataFrame([[
+            transaction_type_encoded, location_encoded, channel_encoded, total_seconds,  
+            transaction_duration, login_attempts, day_of_week, transaction_amount  
+        ]], columns=[
+            "TransactionType", "Location", "Channel", "Time", 
+            "TransactionDuration", "LoginAttempts", "DayOfWeek", "TransactionAmount"
+        ])
+
+        if model_choice == "Isolation Forest":
+            prediction = iso_forest.predict(input_data)[0]
+            anomaly_label = "Anomalous" if prediction == -1 else "Normal"
+        
+        elif model_choice == "One-Class SVM":
+            prediction = one_class_svm.predict(input_data)[0]
+            anomaly_label = "Anomalous" if prediction == -1 else "Normal"
+        
+        elif model_choice == "Local Outlier Factor":
+            # Get the distance of input_data from the neighbors
+            distances, _ = lof_model.kneighbors(input_data)
+            avg_distance = np.mean(distances)
+            
+            # Compare with the LOF threshold
+            anomaly_label = "Anomalous" if avg_distance > lof_threshold else "Normal"
+        
+        # Display result
+        st.write(f"### The transaction is: **{anomaly_label}**")
+
+
+# ---------------- MODEL EVALUATION TAB ---------------- #
+with model_eval:
+    st.header("📊 Model Performance Metrics")
     
-    # Convert time to total seconds since midnight
-    total_seconds = time.hour * 3600 + time.minute * 60
+    # Sidebar to choose which model's evaluation to display
+    eval_model_choice = st.sidebar.radio("Select Model for Evaluation", ["Isolation Forest", "One-Class SVM", "Local Outlier Factor"])
     
-    # Convert categorical values to numeric
-    location_encoded = location_mapping.get(location, -1)  # Default to -1 if not found
-    transaction_type_encoded = transaction_type_mapping[transaction_type]
-    channel_encoded = channel_mapping[channel]
+    # Display evaluation metrics based on selected model
+    if eval_model_choice == "Isolation Forest":
+        st.image("Anomaly_IF_counts.png", caption="Anomaly Counts - Isolation Forest", use_column_width=True)
+        st.write("### 📌 Isolation Forest Performance")
+        st.write("- Detects anomalies based on random sub-sampling of data.")
+        st.write("- **Lower False Positives** in structured transaction data.")
     
-    # Ensure the order of features matches training
-    input_data = pd.DataFrame([[
-        transaction_type_encoded, location_encoded, channel_encoded, total_seconds,  
-        transaction_duration, login_attempts, day_of_week, transaction_amount  # <-- Corrected order
-    ]], columns=[
-        "TransactionType", "Location", "Channel", "Time", 
-        "TransactionDuration", "LoginAttempts", "DayOfWeek", "TransactionAmount"  # <-- Corrected order
-    ])
-
-    # Predict anomaly
-    prediction = iso_forest.predict(input_data)[0]
-    anomaly_label = "Anomalous" if prediction == -1 else "Normal"
+    elif eval_model_choice == "One-Class SVM":
+        st.image("Anomaly_OCSVM_counts.png", caption="Anomaly Counts - One-Class SVM", use_column_width=True)
+        st.write("### 📌 One-Class SVM Performance")
+        st.write("- Uses a hyperplane to separate normal from anomalous data.")
+        st.write("- **Better suited for small datasets** but may be computationally expensive.")
     
-    # Display result
-    st.write(f"### The transaction is: **{anomaly_label}**")
+    elif eval_model_choice == "Local Outlier Factor":
+        st.image("Anomaly_LOF_counts.png", caption="Anomaly Counts - Local Outlier Factor", use_column_width=True)
+        st.write("### 📌 Local Outlier Factor (LOF) Performance")
+        st.write("- Uses density-based analysis to detect anomalies.")
+        st.write("- **Best for identifying local anomalies**, but requires careful tuning of `k-neighbors`.")
+    
+    st.image("silhouette_scores.png", caption="Silhouette Scores for All Models", use_column_width=True)
+    
+    st.header("Comparison")
+    st.write("OCSVM (One-Class SVM) is likely overfitting or being too aggressive in marking transactions as anomalies. Its silhouette score of 0.00 suggests poor cluster structure, meaning it does not effectively separate normal vs. anomalous transactions. LOF (Local Outlier Factor) is the best-performing model because it has the highest silhouette score (0.16), indicating it maintains a clear distinction between anomalies and normal transactions. Isolation Forest (IF) is a close second, performing slightly worse than LOF but still better than OCSVM.")
+
+

local_outlier_factor_model.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:55b4d234fd78b07e802b61b9bb3068e689292f69d3c01e720e89d2c659eb2cdd
+size 806085

lof_threshold.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2368780ca28a9edb8eddf0fd571b650efa2623d358a7a0e183f1ee12a16945c7
+size 116

one_class_svm_model.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0d2d2ac14a8d63c7718450c87429541c614e87f9e4bba136596c3a0f438b49e
+size 212863