Multi-Input Deep Learning for Hotel Review Score Prediction

Introduction

This project addresses the challenge of predicting individual customer satisfaction scores based on a large-scale, heterogeneous dataset of hotel reviews. The goal was to develop an end-to-end deep learning pipeline that moves beyond coarse sentiment classification to predict continuous reviewer scores (regression task). By integrating unstructured text with structured metadata, the project demonstrates how deep learning can drive business intelligence in the hospitality industry.

Methodology

The solution utilizes a multi-input architecture to process non-homogeneous data sources:

1. Data Preprocessing

• Text Cleaning: Merged positive and negative review fields and removed placeholders like "No Negative".
• Tokenization & Padding: Applied sequence padding to a fixed length of 200 tokens.
• Numerical Scaling: Used Min-Max Scaling on hotel metadata (latitude, longitude, average score) for numerical stability during gradient descent.

2. Model Architectures

I implemented and compared three distinct deep learning architectures to evaluate their text processing capabilities:

• Convolutional Neural Networks (CNN): Used 1D-convolutional filters to extract local n-gram features.
• Long Short-Term Memory (LSTM): Designed to capture long-range sequential dependencies and sentiment flow.
• Gated Recurrent Units (GRU): An optimized recurrent architecture for computational efficiency and sequential memory.

3. Optimization Strategy

Models were optimized using the Hyperband algorithm via Keras Tuner. Key tuned hyperparameters included embedding dimensions, recurrent units/filters, dropout rates, and learning rates.

Results

The experimental results proved that recurrent architectures (specifically GRU) outperformed convolutional models in capturing the nuances of subjective review text.

Model (Tuned)	MSE	RMSE	R² (Explained Variance)
CNN	1.1039	1.0507	0.5856
LSTM	1.0641	1.0315	0.6006
GRU	1.0447	1.0221	0.6079

Key Finding: The Tuned GRU model achieved the highest accuracy, explaining 60.8% of the variance in reviewer scores. This suggests that the simplified gating mechanism of the GRU provides an ideal balance of learning capacity and generalization for large-scale text regression.

Dataset Note

• File Size: ~45.6 MB (515,000+ rows)
• Storage: In accordance with professional data engineering best practices, the raw dataset is omitted from this repository due to its large size. It should be downloaded from its original source and stored locally in the /data directory.

How to Run

1. Environment Setup:

   pip install tensorflow pandas scikit-learn matplotlib keras-tuner

2. Dataset Placement: Place the raw Booking.com hotel reviews dataset inside a folder named data/ in your project root directory.
3. Execution: Run the preprocessing pipeline, followed by model_tuning.py to initiate the Hyperband search, and use evaluate_model.py to verify the final performance metrics.