Conference Report: Recommender Systems in Tourism

Antonio Moreno, Laura Sebastiá, Pieter VansteenwegenIEEE Intell. Informatics Bull 2015 · OpenReview

Introduction

This report provides a high-level overview of the TouRS workshop held at the 2015 ACM Conference on Recommender Systems (RecSys). It summarizes the state-of-the-art in Tourism Recommender Systems (TRS), focusing on how researchers are moving beyond simple "item-to-user" matching to solve the complex, multi-dimensional challenges inherent in travel planning.

The Complexity of Travel Planning

Planning a trip is significantly more complex than choosing a movie or a book. Tourism involves high-stakes decisions, multi-day itineraries, and, most importantly, it is rarely a solo activity. The workshop identified several critical gaps in traditional recommender systems when applied to the tourism domain:

  1. Group Dynamics: Recommendations must satisfy families or groups of friends who often have conflicting preferences.
  2. Itinerary Planning: Suggestions cannot exist in a vacuum; they must be ordered logically based on geography, opening hours, and physical constraints such as walking speed and fatigue.
  3. Semantic Context: Standard collaborative filtering often fails to capture the underlying "meaning" of a location. Systems require a deeper understanding of the relationships between categories (e.g., the nuance between "Art" and "Culture") to provide diverse and relevant suggestions.

Methodological Pillars

The workshop's contributions were categorized into four primary methodological approaches:

Group Recommendation Strategies

Two primary strategies for group decision-making were discussed. The first involves merging individual recommendation lists using an "average without misery" strategy, which prioritizes avoiding options that any single group member strongly dislikes. The second utilizes an iterative consensus-reaching process, where individual user profiles are dynamically updated through interaction until the group reaches a high degree of agreement.

Genetic Algorithms for Itinerary Construction

For systems like CT-Planner5, genetic algorithms are employed to construct complex routes. These algorithms respect user-defined constraints such as visit duration, walking difficulty, and specific thematic interests (e.g., prioritizing nature over shopping), evolving a schedule that optimizes for the user's unique constraints.

Semantic Ontologies

Several systems, including SigTur and Troovel, leverage ontologies to map relationships between different activity types. One novel approach defines "semantic similarity" as the ratio between shared ancestors and total ancestors within a concept hierarchy, allowing the system to understand the conceptual proximity of different tourist attractions.

Diversity-Aware k-NN

A significant theoretical contribution explored the "diversity prediction theorem." This suggests that incorporating user profiles that are dissimilar to the target user can actually improve recommendation performance. By looking at a broader "wisdom of the crowd" that includes diverse perspectives, systems can reduce global error and provide more robust results.

Data Sources and Case Studies

While the workshop did not introduce a single benchmark dataset, it highlighted various data sources used to validate these theories:

  • TripAdvisor: Utilized by the Troovel system to compute relationships between items and categories based on aggregate user ratings.
  • DBpedia: Used as a source for topic categories to represent user profiles and domain items in mobile guides.
  • NFC Interaction Data: A case study at Canterbury Cathedral analyzed data from tourists interacting with physical "smart posters" via Near Field Communication (NFC).
  • Geographical Implementations: Specific real-world tests were focused on the city of Jaén (Spain) and the Mediterranean regions of Costa Daurada and Terres de l’Ebre.

Key Insights and Results

  • The Power of Diversity: The finding that recommender performance can improve when considering dissimilar profiles challenges the traditional "find people exactly like you" approach. This suggests that diversity is not just a secondary goal but a driver of accuracy.
  • Semantic Personalization: The SigTur recommender demonstrated that semantic clustering significantly increases the diversity of recommendations without sacrificing accuracy or computational efficiency. Similarly, using shortest-path semantic distance in a DBpedia-based graph proved effective for personalizing museum tours.
  • Iterative Refinement: Planning is rarely a one-shot calculation. Successful systems treat itinerary building as a collaborative, iterative process between the human and the AI, allowing for real-time adjustments.

Future Directions

The workshop concluded by identifying broad challenges for the next generation of Tourism Recommender Systems:

  • Dynamic Preference Management: There is a pressing need for systems that manage shifting user preferences in real-time as users interact with their environment via mobile devices.
  • Physical Context Integration: Future systems must better integrate physical factors—such as weather, current location, and physical fatigue—into the recommendation loop.
  • Industry Adoption: While academic research into semantic and group-based techniques is advanced, a gap remains between these innovations and their full adoption by major commercial travel platforms.

Summary updated

BibTeX
@inproceedings{moreno2015conference,
  author = {Antonio Moreno and Laura Sebastiá and Pieter Vansteenwegen},
  title = {{Conference Report: Recommender Systems in Tourism}},
  booktitle = {IEEE Intell. Informatics Bull},
  year = {2015},
  url = {https://openreview.net/forum?id=GKy6cmjh14},
}
Figure