Abstract
Traditional Point of Interest (POI) recommender systems often fail to account for the dynamic nature of user preferences, leading to repetitive suggestions. This paper introduces Ricochet, a novel system that utilizes Semantic Web technologies and the concept of "complementarity" to balance a user's daily rhythm. By classifying POIs based on "intensities of expressiveness," Ricochet recommends activities that provide a transition—such as a quiet park following a high-intensity museum visit—rather than just more of the same.
Introduction
Current POI recommenders, such as those integrated into Yelp or Foursquare, primarily rely on collaborative filtering or basic contextual data like distance and weather. A significant limitation of these systems is their inability to recognize the sequential relationship between activities. If a user enjoys cultural sites, a standard system might suggest multiple museums in a row, ignoring the physical or mental fatigue that follows a high-intensity activity. Ricochet addresses this by modeling POIs as catalysts for specific "feelings" and seeking to satisfy the resulting needs through complementary transitions.
Methodology
The Ricochet system is built upon three core components: a qualitative user study identifying recommendation criteria, the OntoPOI ontology, and a complementarity-based deduction logic.
The Complementarity Principle
The central innovation is the "Complementarity Rule." The system posits that a POI (POI 1) causes a certain "feeling" or state in the user, which is best addressed by a subsequent POI (POI 2) that satisfies that state.
System Architecture
Ricochet functions as a semantic re-ranking layer over existing data. It retrieves POIs via the Yelp API, translates them into RDF triples using the OntoPOI ontology, and processes them through the Java Jena API. POIs are categorized into four intensity levels (High to Low). The recommendation engine calculates a final score where complementarity is weighted three times more heavily than weather or time of day.
Key Insights
- Rhythm of Daily Life: Effective recommendations should facilitate transitions between different energy levels.
- Semantic Filtering: By adding a layer of semantic understanding to raw location data, the system can provide more "human-like" suggestions.
- POI as a Catalyst: Moving beyond categories, the system views locations through the lens of the physiological or psychological states they induce.
Evaluation and Results
The researchers compared the complementarity-aware system (R1) against a standard context-aware version (R2) in a study involving 10 participants in Paris.
- Performance Metrics: R1 demonstrated superior precision and modified recall compared to R2.
- Ranking Quality: Using nDCG (normalized Discounted Cumulative Gain), R1 provided significantly better-ranked suggestions.
- User Experience: Qualitative feedback indicated that R1 felt more intuitive. For instance, after a high-intensity cinema visit, R1 suggested a low-intensity juice bar, whereas R2 suggested another cinema.
Conclusion and Future Work
Ricochet demonstrates the value of incorporating semantic complementarity into POI recommendations. While currently limited by a small sample size and dependency on external APIs, future iterations could integrate real-time biometric data from mobile sensors (e.g., heart rate, acceleration) to more accurately measure user fatigue and refine intensity parameters.
Summary updated