Abstract
POI recommendation is a key task in tourism information systems. However, in contrast to conventional point of interest (POI) recommender systems, the available data is extremely sparse; most tourists visit a few sightseeing spots once and most of these spots have no check-in data from new tourists. Most conventional systems rank sightseeing spots based on their popularity, reputations, and category-based similarities with users' preferences. They do not clarify what users can experience in these spots, which makes it difficult to meet diverse tourism needs. To this end, in this work, we propose a mechanism to recommend POIs to tourists. Our mechanism includes two components: one is a probabilistic model that reveals the user behaviors in tourism; the other is a pseudo rating mechanism to handle the cold-start issue in POIs recommendations. We carried out extensive experiments with two datasets collected from Flickr. The experimental results demonstrate that our methods are superior to the state-of-the-art methods in both the recommendation performances (precision, recall and F-measure) and fairness. The experimental results also validate the robustness of the proposed methods, i.e., our methods can handle well the issue of data sparsity.
Introduction
Most Point-of-Interest (POI) recommendation systems are designed for routine daily life, such as predicting coffee shop visits based on established habits. Tourism, however, is fundamentally non-routine and adventurous, leading to extreme data sparsity. Most tourists visit a destination only once, see a few spots, and never return.
Existing systems often suffer from two primary issues:
- Popularity Bias: They tend to recommend the same top-tier landmarks to all users, ignoring niche interests and failing to provide fair exposure to lesser-known spots.
- The Cold-Start Problem: Systems often have zero check-in history for new tourists in a city, making personalized recommendations difficult.
This research argues for modeling the semantic experience (e.g., "eating," "playing," "shopping") rather than just geographic coordinates to better understand tourist intent.
Methodology
The proposed framework centers on the Basic User Experience Model (B-UEM), a probabilistic generative model that shifts the focus from "where to visit" to "what to experience."
The Generative Approach
Unlike standard models treating users and locations as independent, B-UEM introduces a latent factor: the User Experience ($g$). The model assumes an "experience" generates four observed variables: the tourist ($u$), activities/words ($w$), the specific POI ($l$), and the time slot ($t$).
By utilizing this hidden layer, the model identifies patterns despite data sparsity. For instance, it can learn that a specific "experience" involves "Zen meditation" at certain temples, linking users with similar interests even if they have no shared physical visit history.
Pseudo-Rating Mechanism
To address the cold-start issue, a "pseudo-rating" system was developed. New users rate a small set of images or keywords, which the system then maps back to the latent "experience" factors using Bayesian inference. This allows for immediate, personalized recommendations based on the learned behaviors of previous tourists.
Experimental Evaluation
Datasets
The study utilized two real-world datasets from Flickr, leveraging the rich semantic data in photo tags:
- YFCC100M Kyoto: 783 tourists and ~1,500 spots in Kyoto.
- User Visits: A multi-city dataset covering 8 global cities (e.g., Budapest, Delhi, Edinburgh) to test cross-urban robustness.
Key Results and Insights
- Superior Accuracy: In "time-split" tests, the proposed models significantly outperformed baselines like GeoMF++ and SAE-NAD. B-UEM-al achieved an F@5 score of ~0.201, compared to GeoMF's 0.068.
- Cold-Start Success: The pseudo-rating mechanism maintained high performance for new users where traditional Matrix Factorization methods typically fail.
- Fairness and Diversity: Using the Gini Index, the authors demonstrated that their model distributes recommendations across a wider variety of POIs, reducing popularity bias and overcrowding.
- Contextual Influence: Variations like ST-UEM and S-UEM were tested to see if location or time influenced behavior. While location does impact activity, the basic B-UEM was often sufficient, suggesting the "Experience" factor captures much of this context inherently.
Conclusion
The B-UEM framework successfully addresses data sparsity and cold-start issues in tourism recommendation by centering on user experience. While effective, the model's reliance on automated tag extraction suggests that improvements in scene recognition would further enhance performance. Future work will focus on optimizing the computational efficiency of the pseudo-rating mechanism to support real-time mobile applications.
Summary updated