New York Airbnb Accommodation Recommender

This project provides a recommendation system for New York Airbnb accommodations using combined features such as text descriptions, images, price, and location data.

The dataset used for this project was collected from Inside Airbnb.

📥 Dataset

The full prepared dataset can be accessed using this link.

🚀 How to Use

Before running the application, ensure the KNN model and combined features are saved. You can generate these by running the notebook:

save_compute_similarities.ipynb

To launch the Streamlit app, execute the following command:

streamlit run airbnb_knn.py

🛠️ Implementation Pipeline

The project implementation follows the pipeline below:

🎯 Project Goals

• Build a robust recommender system combining multiple features (price, description, images, location).
• Utilize advanced text analysis (TF-IDF, Sentiment Analysis).
• Compute similarity between listings using Cosine Similarity and KNN.

Key Methods

1. TF-IDF (Term-Frequency, Inverse Document Frequency Vectorizer)
   • Quantifies the importance of a term in a document relative to its frequency and rarity across multiple documents.
     • Term Frequency (TF): Relative frequency of a term in a document.
     • Inverse Document Frequency (IDF): Measures how rare a term is across all documents.
2. Sentiment Analysis
   • Sentiment analysis is performed using VADER (Valence Aware Dictionary and sEntiment Reasoner).
   • Reference: Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text.
   • VADER generates a Compound Score: a normalized sum of lexicon ratings ranging from -1 (most negative) to +1 (most positive).
3. Image Features
   • Image Histograms (HSV): Histograms of Hue, Saturation, and Value capture the pixel value distribution in images.
   • Similarity: Image similarity is computed by comparing HSV histograms.
4. Cosine Similarity
   • Cosine similarity is calculated between textual descriptions and image HSV vectors to measure feature similarity.
5. K-Nearest Neighbors (KNN)
   • Recommendations are generated using the KNN algorithm with:
     • k = 10
     • Combined features: price, location, image similarity, description similarity, and polarity.