Airbnb Hybrid Recommendation System

A multi-stage recommendation engine that takes natural language queries (e.g., "quiet cabin in the woods with a hot tub under $200/night") and returns the best-matching Airbnb listings, ranked using a pipeline that combines semantic search, graph algorithms, review analysis, and multi-criteria decision making.

Built on real data: over 10,000 listings and 500K+ reviews from Inside Airbnb.

How It Works

User Query: "quiet cabin with hot tub under $200"
    │
    ▼
┌─────────────────────────────────────────────────┐
│  STAGE 1: VECTOR RETRIEVAL                       │
│  Embed query + listings with MiniLM              │
│  Cosine similarity → Top 50 candidates           │
├─────────────────────────────────────────────────┤
│  STAGE 2: GRAPH + PERSONALIZED PAGERANK          │
│  Item-item graph (Jaccard on amenity features)   │
│  + User-item bipartite graph (from reviews)      │
│  → Authority scores for candidates               │
├─────────────────────────────────────────────────┤
│  STAGE 3: REVIEW RELEVANCE (Cross-Encoder)       │
│  Score candidate reviews against user query      │
│  → "Do past guests describe what user wants?"    │
├─────────────────────────────────────────────────┤
│  STAGE 4: TOPSIS MULTI-CRITERIA RANKING          │
│  Combine: cosine_sim, pagerank, wilson_score,    │
│  review_relevance, price → closeness score [0,1] │
│  → Final top-K recommendations                   │
└─────────────────────────────────────────────────┘

Key Algorithms

Algorithm	Purpose
Cosine Similarity (on sentence embeddings)	Semantic matching — "quiet" ≈ "peaceful"
Jaccard Similarity (on binary amenity features)	Graph edge construction — which listings offer similar amenities
Personalized PageRank	Authority — boost listings connected to query-relevant candidates
Wilson Score CI	Rating confidence — penalizes low-review-count listings
Cross-Encoder (ms-marco-MiniLM)	Review relevance — do reviews match what the user wants?
TOPSIS	Final ranking — optimal compromise across all criteria

Project Structure

src/
├── data/
│   ├── loader.py              # Raw CSV → clean parquet
│   └── features.py            # Amenity parsing (25 binary features) + Wilson score
├── retrieval/
│   ├── embeddings.py          # Generate text embeddings (all-MiniLM-L6-v2)
│   └── vector_search.py       # Cosine similarity → top 50 candidates
├── graph/
│   ├── build_graph.py         # Item-item (Jaccard) + user-item (reviews) graph
│   └── pagerank.py            # Personalized PageRank scoring
├── scoring/
│   ├── cross_encoder.py       # Query vs review text relevance
│   └── wilson.py              # Rating confidence interval
├── ranking/
│   └── topsis.py              # Multi-criteria final ranking
└── pipeline.py                # End-to-end orchestrator

Data

• Listings: 4,199 properties with 90 columns → cleaned to 3,635 rows with 25 engineered amenity features
• Reviews: 506,132 guest reviews with free-text comments
• Source: Inside Airbnb (Portland, OR)

Setup

pip install -r requirements.txt

Run preprocessing (once):

python -m src.data.loader
python -m src.retrieval.embeddings
python -m src.graph.build_graph

Run a query:

python -m src.pipeline "quiet cabin in the woods with a hot tub"

Tech Stack

Component	Tool
Embeddings	sentence-transformers (all-MiniLM-L6-v2)
Cross-Encoder	ms-marco-MiniLM-L-6-v2
Graph	NetworkX
Language	Python
Data Processing	pandas, numpy

Design Decisions

• Jaccard over cosine for amenity similarity: Amenity features are binary (yes/no). Jaccard is designed for set comparison and ignores shared absences.
• Personalized PageRank over global: Global PageRank just finds popular listings. PPR seeds on query-relevant candidates so authority scores are query-aware.
• Why Cross-encoder: Cross-encoder answers "do reviews describe what the user actually wants?" — much more useful signal.
• TOPSIS over weighted sum: TOPSIS penalizes being terrible on any single dimension. A listing that's consistently good across all criteria beats one that's amazing on one but bad on others.
• Wilson score over raw ratings: A 5.0★ from 2 reviews is less trustworthy than 4.7★ from 200 reviews. Wilson gives a confidence-adjusted lower bound.