🏠 Hybrid House Price Prediction (Tabular + Images)

A multi-modal ML system combining engineered tabular features with CNN image embeddings

This project builds a hybrid machine learning model that predicts house prices by combining:

• Structured tabular data (bedrooms, bathrooms, square footage, city…)
• Visual information extracted from exterior house photos using EfficientNet
• Gradient-boosted decision trees (XGBoost) for final price prediction

The approach outperforms tabular-only models by capturing implicit visual attributes such as condition, curb appeal, architectural style, landscaping, and exterior quality—features that are typically unavailable or too subjective for users to input manually.

---

📂 Dataset

The project uses the public Kaggle dataset:

🔗 House Prices and Images - SoCal

It contains:

• 15,474 Southern California listings
• Tabular metadata (bed/bath/sqft, city, price, etc.)
• One exterior house image per listing
• Clean structure with no missing values

---

✨ Project Highlights

✔️ Advanced Feature Engineering

• log₁₊price target transformation
• Spaciousness metrics: sqft_per_bed, sqft_per_bath
• Total rooms
• Log-transformed sqft
• Target-encoded city (mean log-price per city in training split)
• Standardized numeric features

✔️ CNN Transfer Learning (EfficientNet)

• EfficientNetB0 and EfficientNetB3 pretrained on ImageNet
• Fine-tuned EfficientNetB0 (last 10 layers unfrozen)
• Strong augmentation pipeline (flip, zoom, rotation, translation, contrast)
• Extracted 128-dimensional visual embeddings representing house condition & style

✔️ Hybrid Architecture

The hybrid model concatenates:

[scaled tabular features] + [128-dim image embedding]

and feeds the combined vector into a tuned XGBoost regressor.

✔️ Performance

Model	R² (test)
Linear Regression (baseline)	~0.40
Random Forest	~0.42
XGBoost (baseline)	~0.69
XGBoost (feat-engineered + tuned)	~0.78
Hybrid Model (Tabular + Images)	~0.80

---

📁 Repository Structure

📦 FinalProject/
│
├── .gitignore                    # Git ignore rules
├── README.md                     # Project documentation
├── main.ipynb                    # Full training pipeline: tabular, CNN, hybrid
├── Presentation.pdf              # Project presentation slides
│
└── streamlit_app/
    ├── app.py                    # Streamlit demo application
    ├── requirements.txt          # Python dependencies
    └── models/
        ├── city_target_enc.json           # Target-encoding mapping for city feature
        ├── config.json                    # Model configuration (image size, features, etc.)
        ├── effnetb0_simple_lastlp.keras   # EfficientNetB0 variant
        ├── effnetb0_t1_best.keras         # Fine-tuned EfficientNetB0 (best checkpoint)
        ├── effnetb3_t1_best.keras         # Fine-tuned EfficientNetB3 (experimental)
        ├── hybrid_tuned_model.pkl         # Final hybrid regressor (tabular + images)
        ├── hybrid_xgb_model.pkl           # Hybrid XGBoost model
        └── xgb_fe_tuned_pipeline.pkl      # Tabular-only preprocessing + XGBoost

---

🧠 Modeling Pipeline

1️⃣ Exploratory Data Analysis

• Verified dataset integrity (15,474 rows, all images present)
• Visualized price distribution
• Modeled log-price due to heavy skew
• Analyzed correlations and feature relationships
• Identified location as a major driver → target encoding

---

2️⃣ Tabular Modeling

Baseline models:

• Linear Regression
• Random Forest
• XGBoost

XGBoost clearly outperformed with R² ≈ 0.69, but not enough → needed better features.

Feature engineering dramatically improved results:

• Spaciousness ratios
• log_sqft
• total_rooms
• Target encoding

XGBoost FE + tuning ⇒ R² ≈ 0.78

---

3️⃣ Image Modeling (CNN)

Models tested:

• EfficientNetB0 (baseline)
• EfficientNetB3
• EfficientNetB0 fine-tuned (last 10 layers unfrozen)

CNNs alone performed poorly for price prediction, but:

EfficientNetB0 tuned produced the best embeddings, used in the hybrid model.

Images capture:

• Renovation quality
• Architectural style
• Curb appeal
• Landscaping
• General exterior condition

---

4️⃣ Hybrid Model

Architecture:

1. Encode tabular data → StandardScaler
2. Compute image embedding → EfficientNetB0 tuned
3. Concatenate → [tabular_scaled | embedding_128]
4. Predict log-price → XGBoost regressor
5. Convert back to USD using expm1

Performance:

• R² ≈ 0.80 on test set
• Visual features successfully improved the model

---

🚀 Streamlit Demo

A Streamlit app is included so users can:

• Input property details
• Optionally upload an exterior photo
• Choose:
  • Tabular-only price estimate
  • Hybrid model price estimate
• Get a final predicted price + explanation text

---

⚙️ Installation & Setup

Prerequisites

• Python 3.8 or higher
• pip package manager

1️⃣ Clone the repository

git clone https://github.com/your-username/house-price-prediction.git
cd house-price-prediction/streamlit_app

2️⃣ Create a virtual environment

# Windows
python -m venv .venv
.\.venv\Scripts\activate

# Mac/Linux
python3 -m venv .venv
source .venv/bin/activate

3️⃣ Install dependencies

pip install -r requirements.txt

4️⃣ Run the Streamlit app

streamlit run app.py

5️⃣ Use the interface

• Fill in tabular features
• Upload an image (optional)
• Choose model type
• View the predicted price

---

📌 Model Artifacts

Inside streamlit_app/models:

File	Purpose
xgb_fe_tuned_pipeline.pkl	Full preprocessing + tuned XGBoost (tabular-only model)
effnetb0_t1_best.keras	Fine-tuned EfficientNetB0 - main CNN for image embeddings
effnetb0_simple_lastlp.keras	EfficientNetB0 variant (alternative architecture)
effnetb3_t1_best.keras	Fine-tuned EfficientNetB3 (experimental, higher capacity)
hybrid_xgb_model.pkl	Hybrid regressor (tabular + image features)
hybrid_tuned_model.pkl	Final tuned hybrid model
city_target_enc.json	Target-encoding mapping for city feature
config.json	Model configuration (image dimensions, feature list)

Note: The app automatically loads the best-performing model configuration.

---

📊 Presentation

A detailed project presentation is available in Presentation.pdf, covering:

• Problem statement and motivation
• Data exploration and insights
• Modeling approach and architecture
• Results and performance metrics
• Demo and future improvements

---

🧭 Future Work

• Incorporate multiple images per listing (interior + exterior)
• Use satellite imagery to capture neighborhood quality
• Explore multi-task learning (predict price + condition score)
• Implement SHAP for full model explainability
• Deploy the system as a full API + web app

---

👤 Author

Javier García Esteve — Final project for the IRONHACK Data Science Bootcamp.

Includes end-to-end ML engineering, deep learning, image modeling, and application deployment.

---

📄 License

This project is open source and available under the MIT License.