House Rent Prediction Using Linear Regression

An end-to-end machine learning regression project that predicts house rent based on property features such as size, BHK count, city, furnishing status, and tenant preferences.

---

Table of Contents

• Project Overview
• Dataset
• Tech Stack
• Project Pipeline
• Exploratory Data Analysis
• Data Preprocessing
• Model Building and Training
• Model Evaluation
• Model Saving and Inference
• Project Structure
• How to Run
• Key Design Decisions

---

Project Overview

This project implements a complete supervised machine learning pipeline to predict monthly house rent in Indian cities. Linear Regression is used as the core algorithm, with careful attention to regression assumptions — including linearity, multicollinearity, and residual autocorrelation — throughout the workflow.

The target variable (Rent) is log-transformed before training to handle skewness and improve the linear relationship with input features. Predictions are reverse-transformed using expm1 to return rent values in the original scale.

---

Dataset

File: House_Rent_Dataset.csv

The dataset contains listings of rental properties across multiple Indian cities with attributes describing the property and rental terms.

Features Used After Preprocessing

Feature	Type	Description
BHK	Numerical	Number of Bedrooms, Hall, and Kitchen
Size	Numerical	Property size in square feet
City	Categorical	City where the property is located
Area Type	Categorical	Super Area, Carpet Area, or Built Area
Furnishing Status	Categorical	Furnished, Semi-Furnished, or Unfurnished
Tenant Preferred	Categorical	Bachelors, Family, or Bachelors/Family
Point of Contact	Categorical	Contact Owner, Contact Agent, or Contact Builder

Columns Dropped

Column	Reason for Removal
Floor	High cardinality, inconsistent formatting
Area Locality	Too granular; leads to sparse encoding
Posted On	Date field not relevant to rent price modeling
Bathroom	Dropped after VIF analysis revealed multicollinearity with BHK and Size

Target Variable

• Raw: Rent (monthly rent in INR)
• Transformed: log_rent = log(Rent) — used for model training
• Prediction output: Reverse-transformed via np.expm1(log_rent)

---

Tech Stack

Library	Purpose
pandas	Data loading, manipulation, and feature engineering
numpy	Numerical operations and log transformation
matplotlib	Static visualizations
seaborn	Statistical plots and styled visualizations
scikit-learn	Train-test split, scaling, model training, metrics
statsmodels	VIF computation, Durbin-Watson test
joblib	Saving and loading trained model and scaler

---

Project Pipeline

Raw CSV Data
     |
     v
Data Loading & Inspection
     |
     v
Exploratory Data Analysis (EDA)
     |
     v
Outlier Removal
     |
     v
Log Transformation of Target
     |
     v
Feature Selection (VIF Check)
     |
     v
One-Hot Encoding of Categorical Features
     |
     v
Train-Test Split (80/20)
     |
     v
Feature Scaling (StandardScaler)
     |
     v
Linear Regression Model Training
     |
     v
Model Evaluation (R2, MAE, RMSE, Residual Analysis)
     |
     v
Reverse Transformation & Prediction
     |
     v
Model Serialization (joblib)

---

Exploratory Data Analysis

The EDA phase examines both the distribution of individual features and their relationship with the target variable.

Numerical distributions — Histograms with KDE curves are plotted for Rent, Size, BHK, and Bathroom to understand spread and skewness.

Categorical distributions — Count plots are generated for City, Area Type, Furnishing Status, Tenant Preferred, and Point of Contact.

Rent vs numerical features — Scatter plots of Rent against Size, BHK, and Bathroom are used to assess linearity before and after log transformation.

Rent vs categorical features — Box plots show how rent varies across categories, helping identify which categorical groups have the most pricing power.

Correlation heatmap — A heatmap of the numeric feature correlation matrix is used to detect pairwise relationships and potential multicollinearity.

---

Data Preprocessing

Outlier Removal

Outliers are capped at the 99th percentile separately for Rent and Size to reduce the influence of extreme values without discarding too much data.

rent_upper_limit = df["Rent"].quantile(0.99)
df = df[df["Rent"] < rent_upper_limit]

size_upper_limit = df["Size"].quantile(0.99)
df = df[df["Size"] < size_upper_limit]

Log Transformation

The Rent column is right-skewed. Applying a log transformation produces a more normally distributed target and improves the linear fit.

df["log_rent"] = np.log(df["Rent"])

Multicollinearity Check (VIF)

Variance Inflation Factor (VIF) is computed for BHK and Size after Bathroom is removed. Low VIF scores confirm that multicollinearity is not a concern for the remaining numerical features.

One-Hot Encoding

All remaining categorical columns are encoded using pd.get_dummies with drop_first=True to avoid the dummy variable trap, and output as integer columns.

---

Model Building and Training

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

Scaling — StandardScaler is applied to BHK and Size. The scaler is fit only on the training set and then applied to the test set to prevent data leakage.

Model — sklearn.linear_model.LinearRegression is trained on the scaled training data.

---

Model Evaluation

The model is assessed using four approaches:

Regression Metrics

Metric	Description
R2 Score	Proportion of variance in log_rent explained by the model
MAE	Mean absolute error on the log scale
RMSE	Root mean squared error on the log scale

Actual vs Predicted Plot

A scatter plot of actual vs predicted log_rent values, overlaid with a perfect prediction line (y = x), shows how closely predictions track true values.

Residual Analysis

• Residual distribution — A histogram with KDE checks whether residuals are approximately normally distributed and centered at zero.
• Residuals vs predicted — A scatter plot checks for homoscedasticity (constant variance of residuals across predicted values).

Durbin-Watson Test

The Durbin-Watson statistic is computed to test for autocorrelation in residuals. A value close to 2.0 indicates no significant autocorrelation, confirming that the independence assumption holds.

---

Model Saving and Inference

The trained model and scaler are saved for reuse without retraining:

joblib.dump(model, "house_rent_model.pkl")
joblib.dump(scaler, "scaler.pkl")

Sample Prediction

A sample prediction is demonstrated for a 2 BHK, 1200 sq. ft. Semi-Furnished property in Mumbai with the following steps:

1. Create a DataFrame with the raw input features
2. Apply pd.get_dummies and reindex to match training columns
3. Scale numerical features using the saved scaler
4. Predict log_rent using the saved model
5. Reverse-transform: predicted_rent = np.expm1(model.predict(sample_data))

---

Project Structure

├── House_RentPrediction_project.ipynb   # Main notebook with full pipeline
├── House_Rent_Dataset.csv               # Input dataset
├── house_rent_model.pkl                 # Serialized trained Linear Regression model
├── scaler.pkl                           # Serialized StandardScaler
└── README.md                            # Project documentation

---

How to Run

1. Clone or download the repository.

2. Install the required dependencies:

   pip install pandas numpy matplotlib seaborn scikit-learn statsmodels joblib

3. Place House_Rent_Dataset.csv in the same directory as the notebook.

4. Open the notebook:

   jupyter notebook House_RentPrediction_project.ipynb

5. Run all cells in order. The notebook will generate all visualizations, print evaluation metrics, and save the model files to the working directory.

---

Key Design Decisions

Log transformation of target — Raw rent values are heavily right-skewed. Log transformation normalizes the distribution and strengthens the linear relationship between features and the target, which is a core assumption of Linear Regression.

99th percentile outlier capping — Hard removal of extreme outliers in Rent and Size was preferred over IQR-based clipping to retain more data while still limiting the distortion caused by very high-value listings.

Bathroom dropped via VIF — Rather than relying on correlation alone, VIF was used to formally detect multicollinearity. Bathroom showed redundancy with BHK and Size and was removed to improve model stability.

drop_first=True in encoding — Avoids perfect multicollinearity introduced by dummy variables representing all levels of a categorical feature.

Scaler fit only on train set — Prevents data leakage by ensuring that test set statistics do not influence the scaling parameters used during training.