Omerinbar/wanderlust-ai-dataset

✈️ Wanderlust AI — Travel Recommendation Dataset

A synthetic dataset of AI-generated travel destinations and points of interest, built for a two-layer travel recommendation and itinerary-generation application.

---

📓 Project Notebook

The full project notebook (data generation, cleaning, EDA, recommendation system, generation, and application code) is available here: Wanderlust_AI_Final_Project.ipynb

---

📋 Project Overview

Wanderlust AI matches a user's travel "vibe" to real destinations (Layer A), recommends real points of interest within a chosen destination (Layer B), and generates a personalized day-by-day itinerary.

Content generated by	Meta's Llama 3.1-8B-Instant, via Groq's free inference API
Method	AI-generated descriptions of real, named places (no web scraping)
Raw generated dataset	35,536 rows (3,370 destinations + 32,166 POIs)
Final clean dataset	26,117 rows (2,861 destinations + 23,256 POIs)
Coverage	220 real cities across 6 continents, 13 traveler personas

---

📦 Part 1: Synthetic Data Generation

Content was generated using Meta's Llama 3.1-8B-Instant, accessed via Groq's free inference API (an approved course tool). Real place names were used — drawn from the model's own trained knowledge — paired with AI-generated descriptions.

Generation process:

• 220 real cities selected across 6 regions, chosen for being well-documented tourist destinations (reducing hallucination risk)
• Each city described from the perspective of 13 distinct traveler personas
• POI categories dynamically assigned per city based on that city's vibe tags, then generated with real-sounding names and descriptions

Raw output (before any cleaning):

File	Rows
Destinations (raw)	3,370
POIs (raw)	32,166
Combined total	35,536

This exceeds the assignment's 10,000-row minimum by a wide margin, even before cleaning began.

---

🧹 Part 2: Data Cleaning & Exploratory Data Analysis

Cleaning Performed

• Removed 8,176 exact-duplicate POI entries and 728 fuzzy-matched near-duplicates (90% name similarity threshold, matched within the same category to avoid false positives)
• Normalized inconsistent casing and merged near-duplicate labels across personas, POI categories, and budget levels
• Removed 5 malformed entries where the model leaked prompt-instruction text into output fields
• Removed 4 entries with empty descriptions

Dataset Structure

destinations_final.csv — descriptions of 220 real cities, each written from the perspective of 13 personas (budget backpacker, luxury traveler, foodie, family with young kids, solo female traveler, digital nomad, honeymoon couple, adventure/outdoor enthusiast, retiree/senior traveler, group of friends, wellness/yoga retreat seeker, photographer/instagrammer, history/culture buff).

Column	Description
city	City name
country	Country
region	Continent/region
persona	Traveler persona this entry is tailored to
vibe_tags	List of vibe/theme tags
budget_level	shoestring / mid-range / splurge
best_season	Recommended travel season
description	AI-generated, persona-tailored description

pois_final.csv — real, named points of interest within each city.

Column	Description
name	Place name
category	POI category
description	AI-generated description
city	City the POI is located in
country	Country

Exploratory Data Analysis

Dataset Overview

City coverage skews toward Europe (78 cities) and Asia (56 cities), reflecting deliberate selection of well-documented tourist destinations. Personas are evenly represented, though description length varies meaningfully by persona — "history/culture buff" entries run notably longer (median 500 characters) than others (380-420 characters). POI categories are dominated by restaurant, cafe, and hotel — the same categories our quality audit found least reliable (see below).

Region → Budget Level Breakdown

An interactive version of this chart (click to drill down) is available in the project notebook.

Budget-level distribution is fairly balanced within every region, confirming even assignment rather than correlation with a region's real-world cost reputation.

POI Categories by Region — Treemap

An interactive version of this chart is available in the project notebook.

Restaurant and cafe dominate every region. Europe shows the richest category diversity, reflecting its concentration of "romance" and "history"-tagged cities.

POI Coverage by City

POI counts per city range from 44 to 179 (average 105.7). Coverage is driven by how many vibe-tag categories a city qualifies for, not by real-world fame.

Description Length: Destinations vs. POIs

Destination descriptions (mean 418 characters) are substantially longer than POI descriptions (mean 98 characters) — an intentional design choice reflecting their different roles.

POI Category Distribution by Region

Core categories (restaurant, cafe, hotel) appear consistently across every region. Niche categories show designed geographic concentration — "scenic viewpoint" appears only in Europe, "spa/wellness spot" only in Asia and North America — a consequence of our vibe-tag-based category system.

Data Quality Audit

We manually verified 80 randomly sampled entries against real-world sources.

Verdict	Count	Percentage
✅ Accurate	51	63.7%
❌ Wrong/Hallucinated	19	23.75%
⚠️ Unsure	10	12.5%

Statistical Confidence

Accuracy: 63.7% (95% CI: 53.2%–74.3%, n=80) — even in the most conservative interpretation, accuracy sits well above chance level.

Accuracy by POI Category

Category	Accurate	Wrong	Unsure	Error Rate
Restaurant	7	12	3	55%
Cafe	3	4	1	63%
Beach bar	2	1	2	60%
Museum	9	1	2	25%
Food market	2	0	1	33%
Outdoor/nature	2	0	1	33%
Hotel	9	1	0	10%
Historic site	7	0	0	0%
Beach	2	0	0	0%

Key finding: restaurants and cafes are markedly less reliable than other categories, despite being the two largest categories by volume. Historic sites, hotels, and beaches were highly reliable — likely because major landmarks and hotel chains are more heavily represented in the underlying model's training data than individual small restaurants and cafes.

Identified Hallucination Pattern

A specific, repeatable error type emerged: the model occasionally takes a real landmark or location name and incorrectly attaches a fictional business type to it.

• "Restaurant Kapuzinerberg" — Kapuzinerberg is a real hill/landmark in Salzburg, Austria; no such restaurant exists
• "Puri Saren Agung Restaurant" — Puri Saren Agung is a real palace/place of worship in Ubud, Bali, not a restaurant
• "Dar Chouara Museum" — conflates the real Chouara Tannery in Fes with a fictional museum framing

Note on Destination Count

During Part 3 recommendation-system testing, we discovered 509 duplicate (city, persona) rows among the destinations — caused by resume-logic gaps during Part 1's generation retries, where the same combination was occasionally sent to the model twice across sessions. These were identified and removed, bringing the final destination count from 3,370 down to 2,861 (final combined total: 26,117, as stated in the Project Overview above). See Part 3 for the full bug discovery and fix.

---

🔍 Part 3: Recommendation System (Embeddings)

Embedding Model Selection

We compared 3 pre-trained sentence embedding models: all-MiniLM-L6-v2 (384 dimensions), all-mpnet-base-v2 (768 dimensions), and paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions), evaluating encoding speed and semantic quality against representative "vibe" queries.

The multilingual model consistently returned the most accurate results (zero clear mismatches across test queries, vs. both other models incorrectly surfacing Nairobi — a safari destination — for a beach query), with competitive speed. Selected: paraphrase-multilingual-MiniLM-L12-v2.

System Architecture

• Layer A (recommend_destinations) — matches a free-text vibe query to the best-fitting destinations via FAISS cosine similarity search, with optional filtering by persona, budget, and region
• Layer B (recommend_pois) — recommends relevant points of interest within a chosen destination, using category-aware keyword mapping to prevent category bleed-through (e.g., ensuring a "beach" query doesn't surface a shopping mall due to loose semantic similarity)

Bugs Found and Fixed

• Duplicate destination rows — 509 rows with identical (city, persona) combinations but different AI-generated content, caused by resume-logic gaps during generation retries. Fixed via deduplication, re-encoding, and FAISS index rebuild.
• Category bleed-through in Layer B — fixed via keyword-to-category filtering, verified via before/after comparison.

Embedding Space Analysis

To validate our architectural choice of similarity search over hard clustering, we visualized and quantitatively evaluated both embedding spaces.

PCA — Destination Embeddings by Region

No meaningful clustering by region — points from different regions are thoroughly intermixed, indicating semantic similarity doesn't map cleanly onto geography.

t-SNE — Destination Embeddings by Primary Vibe

Using a technique specifically sensitive to local clustering, colors remain intermixed throughout — confirming, via a second independent method, that vibe categories don't form clean, separate zones in the embedding space.

K-Means Clustering — Destination Embeddings

Silhouette scores remain very low (0.042–0.052) across all tested values of k, quantitatively confirming the data does not naturally organize into discrete, well-separated groups.

K-Means Clustering — POI Embeddings

By contrast, POI embeddings show meaningfully stronger clustering structure (silhouette scores 3–4x higher, peaking at 0.171 for k=9), with resulting clusters aligning closely to real POI categories (hotels, restaurants, museums, cafes, bars, beaches) despite category labels never being used as input.

Conclusion: this consistent pattern across three independent techniques — weak structure for destinations, meaningfully stronger structure for POIs — directly supports our decision to rely on similarity-based retrieval rather than hard clustering for both recommendation layers.

---

🗓️ Part 4: Itinerary Generation

Using the same Llama 3.1-8B-Instant model (via Groq), given a chosen destination and its recommended points of interest, the system generates an original day-by-day itinerary — using only real, recommended places, organized into morning/afternoon/evening activities and tailored to the traveler's stated preferences and persona.

Testing: Verified with two distinct test cases (beach vacation for Punta del Este, historic culture for Helsinki). Both correctly used only provided POIs, with no invented locations.

Limitation identified: Helsinki's dataset showed limited historic-site/museum coverage, causing fallback to broader categories — an honest reflection of uneven category depth across cities.

Critical Bug Found and Fixed: POI Embeddings Misalignment

A structured review of how our data-cleaning steps interacted with our embedding pipeline revealed that POI embeddings had become misaligned with the underlying dataset — encoded before two later rounds of row removal (empty descriptions, then leaked meta-commentary text), without ever being re-encoded.

Fix: reset the dataframe index, re-encoded all 23,256 POI embeddings from the current clean data, and rebuilt the FAISS index. Verified by re-running both itinerary demos and confirming meaningfully different (correct) results, plus category-specific sanity checks (Tel Aviv beach bars, Kyoto temples) returning 100% category-consistent results post-fix.

Note on Model Serving Approach

We use Meta's Llama 3.1-8B-Instant — an open-source HuggingFace model — served via Groq's inference API rather than loaded directly via the transformers library. Loading and running an 8-billion-parameter model directly in Google Colab's free tier is impractical for interactive, real-time generation, requiring significant GPU memory and producing noticeably slower response times. Groq serves this same open-source model with dramatically faster inference, making real-time itinerary generation feasible within our free-tier constraints — we consider this equivalent to using the HuggingFace model itself via an optimized serving layer, not a switch to a different, closed-source model.

---

🎨 Part 5: Gradio Application

The dataset powers a full-featured Gradio application, including:

• Multiple search modes (free-text vibe, tag-based, direct city/country lookup, "surprise me")
• Persona, budget, and region filtering
• Live weather, timezone, destination photos, and interactive maps
• AI-generated budget estimates, packing lists, and local phrase cheat-sheets
• A multi-turn conversational AI assistant with memory, using retrieval-augmented generation to ground responses in real dataset content

---

🎁 Bonus Features

• Live real data integration: weather, timezone, currency conversion, geocoding/maps, and destination photos, all fetched live from free public APIs
• Discord bot integration: a fully conversational Discord bot (@mention-based, with multi-turn memory) reusing the same recommendation engine as the in-app chat
• Fine-tuned model: a LoRA-fine-tuned small model (Qwen2.5-0.5B) for structured travel-intent extraction, trained and evaluated on 147 examples

---

⚠️ Known Limitations

• Restaurant and cafe recommendations should be treated with more caution than historic sites, museums, or hotels, given the documented reliability difference
• Some cities have uneven category depth (e.g., thin historic-site/museum coverage in certain cities), which can cause recommendation fallback to broader categories
• Niche POI categories are geographically concentrated by design, not globally uniform
• A small number of near-duplicate POI entries may remain despite deduplication, due to inherent limitations of name-based fuzzy matching

---

🎯 Intended Use

This dataset powers a two-layer travel recommendation system. It is intended for building recommendation systems, testing embedding models, and exploring synthetic data generation — not as a factual, verified travel guide.

---

📄 License

MIT

Author: Omer Inbar & Ohad Danon | Reichman University — Adelson School of Entrepreneurship | Introduction to Data Science | 2026

🤖 AI Usage Disclosure

This project was completed with assistance from Claude (Anthropic) for code debugging, data cleaning, chart design, and README writing. All analysis, decisions, and interpretations are our own.