🤖 Supervised Learning — AIML Module Project

An end-to-end Supervised Learning project spanning two real-world domains: Medical Biomechanics Classification (KNN) and Bank Loan Conversion Prediction (Logistic Regression, SVM, KNN). Covers data preparation, EDA, model building, class balancing, and hyperparameter tuning.

---

📁 Project Structure

medical-banking-classification-ml/
│
├── project_for_supervised_learning.ipynb   # Main Jupyter Notebook
├── project_for_supervised_learning.html    # HTML export of the notebook
│
├── Normal.csv       # Biomechanics data — Normal class   (100 rows)
├── Type_H.csv       # Biomechanics data — Type H class   (60 rows)
├── Type_S.csv       # Biomechanics data — Type S class   (150 rows)
│
├── Data1.csv        # Bank customer demographics         (5000 rows × 8 cols)
├── Data2.csv        # Bank customer financial profile    (5000 rows × 7 cols)
│
└── README.md

---

🗂️ Project Overview

Part	Domain	Algorithm(s)	Marks
Part A	Medical — Biomechanics Classification	KNN	30
Part B	Banking — Loan Conversion Prediction	Logistic Regression, SVM, KNN	30
Total			60

---

🏥 Part A — Medical Biomechanics Classification (30 Marks)

Domain: Medical
Context: Medical Research University X is conducting deep research on patients with masked conditions. Patient details are anonymised; the AI team receives biomechanics measurements to build a condition-prediction model.

Objective: Classify patient condition (Normal / Type_H / Type_S) from six biomechanical features using a KNN classifier.

---

Datasets: Normal.csv, Type_H.csv, Type_S.csv

All three files share the same 7 columns:

Column	Description	Type
P_incidence	Pelvic incidence angle	float64
P_tilt	Pelvic tilt angle	float64
L_angle	Lumbar lordosis angle	float64
S_slope	Sacral slope	float64
P_radius	Pelvic radius	float64
S_Degree	Spondylolisthesis degree	float64
Class	Patient condition label	string

Dataset Sizes & Class Label Inconsistencies

File	Rows	Class Label Variants	Cleaned Label
Normal.csv	100	'Normal', 'Nrmal'	'normal'
Type_H.csv	60	'Type_H', 'type_h'	'type_h'
Type_S.csv	150	'Type_S', 'tp_s'	'type_s'

⚠️ Data Cleaning Note: Each file contains inconsistent spellings/cases in the Class column that must be unified before combining. After cleaning and merging all three files, the final DataFrame shape is (310, 7) — matching the expected checkpoint.

---

Steps & Tasks

1. Data Understanding

• Read all 3 CSVs into separate DataFrames
• Compare shapes, column names, and data types across all three
• Observe Class label variations per file

2. Data Preparation & Exploration

• Unify Class label variants (e.g., 'Nrmal' → 'normal', 'tp_s' → 'type_s')
• Combine into a single DataFrame — expected shape: (310, 7)
• Print 5 random samples, check null values, generate 5-point summary

3. Data Analysis (EDA)

• Heatmap — Correlation between all 6 biomechanical features
• Pairplot — 3 classes differentiated by colour
• Jointplot — P_incidence vs S_slope relationship
• Boxplot — Feature-wise distribution and outlier detection

4. Model Building

• Features (X): 6 biomechanical columns | Target (Y): Class
• Train/test split: 80:20
• Base model: KNN Classifier
• Evaluation: Accuracy, Precision, Recall, F1-Score (train + test)

5. Performance Improvement

• Tune KNN hyperparameters (e.g., n_neighbors, metric, weights)
• Report improvement deltas (e.g., Accuracy: +X%, Precision: +Y%)
• State which parameters contributed most to improvement

---

🏦 Part B — Bank Loan Conversion Prediction (30 Marks)

Domain: Banking & Marketing
Context: Bank X wants to convert its liability customers (depositors) into asset customers (borrowers). The last campaign had a single-digit conversion rate. The goal is to build an ML model that identifies high-probability loan converters for targeted marketing — aiming to double the conversion rate within the same budget.

Objective: Predict whether a customer will take a LoanOnCard (binary classification) using customer demographic and financial data.

---

Datasets: Data1.csv + Data2.csv → Merged on ID

Data1.csv — Customer Demographics (5,000 rows × 8 columns)

Column	Description	Type
ID	Customer ID (merge key)	int64
Age	Customer's approximate age	int64
CustomerSince	Years as bank customer (masked unit)	int64
HighestSpend	Highest single transaction spend (masked unit)	int64
ZipCode	Customer zip code	int64
HiddenScore	Bank's proprietary score (values: 1–4)	int64
MonthlyAverageSpend	Average monthly spend (masked unit)	float64
Level	Bank's proprietary level (values: 1–3)	int64

Data2.csv — Customer Financial Profile (5,000 rows × 7 columns)

Column	Description	Type
ID	Customer ID (merge key)	int64
Mortgage	Mortgage value (masked unit)	int64
Security	Security asset with bank (0/1)	int64
FixedDepositAccount	Has fixed deposit account (0/1)	int64
InternetBanking	Uses internet banking (0/1)	int64
CreditCard	Uses bank credit card (0/1)	int64
LoanOnCard	Target variable — has credit card loan (0/1)	float64

After merging on ID: shape = (5,000 × 14)

⚠️ Class Imbalance: LoanOnCard is heavily imbalanced — 0 (No Loan): 4,500 | 1 (Loan): 480 | NaN: 20 — a ~9.4:1 ratio requiring balancing before model training.

⚠️ Data Type Note: CreditCard, InternetBanking, FixedDepositAccount, Security, Level, HiddenScore are binary/categorical but stored as int64 — must be cast to object before modelling.

---

Steps & Tasks

1. Data Understanding & Preparation

• Read Data1.csv and Data2.csv into separate DataFrames
• Merge on ID → single DataFrame (5,000 × 14)
• Cast CreditCard, InternetBanking, FixedDepositAccount, Security, Level, HiddenScore → object

2. Data Exploration & Analysis

• Visualise LoanOnCard distribution and share insights on class imbalance
• Check and impute missing values (20 nulls in LoanOnCard)
• Detect and impute unexpected values in categorical columns

3. Data Preparation & Model Building

• Features (X): All columns except ID, ZipCode, LoanOnCard | Target (Y): LoanOnCard
• Train/test split: 75:25
• Base model: Logistic Regression
• Evaluate: Accuracy, Precision, Recall, F1, Confusion Matrix
• Balance LoanOnCard to 50:50 using appropriate technique (SMOTE / oversampling / undersampling)
• Retrain Logistic Regression on balanced data — compare metrics before vs. after balancing

4. Performance Improvement

• Train base models for SVM and KNN
• Tune hyperparameters for each (e.g., C, kernel for SVM; n_neighbors, metric for KNN)
• Select and report final model evaluation metrics
• Document improvement from base → final model

---

🛠️ Tech Stack

• Language: Python 3.x
• Notebook: Jupyter Notebook
• Libraries:

pandas            # Data manipulation
numpy             # Numerical operations
matplotlib        # Static plots
seaborn           # Statistical visualisations (heatmap, pairplot, jointplot, boxplot)
scikit-learn      # KNN, Logistic Regression, SVM, train_test_split, metrics
imbalanced-learn  # SMOTE / class balancing

---

⚙️ Setup & Usage

# Clone the repository
git clone https://github.com/<your-username>/supervised-learning-aiml-project.git
cd supervised-learning-aiml-project

# Install dependencies
pip install pandas numpy matplotlib seaborn scikit-learn imbalanced-learn jupyter

# Launch the notebook
jupyter notebook project_for_supervised_learning.ipynb

---

📈 Key Findings (Summary)

Part A — Biomechanics Classification:

• All three source files contain inconsistent Class label spellings that must be unified before merging ('Nrmal', 'tp_s', 'type_h' variants exist)
• Final combined dataset: 310 patients across 3 conditions (Normal: 100, Type_S: 150, Type_H: 60)
• KNN classification on biomechanical features; performance improves with tuned n_neighbors and distance metric

Part B — Loan Conversion Prediction:

• Severe class imbalance: only ~9.6% of customers converted (LoanOnCard = 1), mirroring the single-digit campaign conversion rate
• 20 missing values in target variable LoanOnCard need imputation before modelling
• Balancing the target to 50:50 significantly improves recall for the minority (loan) class — critical for marketing use cases where missing a potential converter is costly
• Final model comparison across Logistic Regression, SVM, and KNN to identify the best predictor for targeted campaigns

---

📋 Submission Checklist

• .ipynb notebook with all code, outputs, and markdown explanations
• .html export of the notebook
• All code cells have visible outputs
• Insights documented for every question
• No plagiarism

---

📄 License

This project was completed as part of the Great Learning AIML Programme coursework.
Educational use only.

---

Made with 🤖 and Python