Trust Based Recommender System for Semantic Web

Punam Bedi, Harmeet Kaur, Sudeep MarwahaIJCAI 2007

Introduction

Traditional recommender systems, largely based on Collaborative Filtering (CF), often operate as "black boxes." While they effectively suggest items based on user similarity, they suffer from two primary drawbacks: a lack of transparency and an inability to handle dynamic knowledge drift. Users typically trust recommendations from known, reliable sources more than anonymous algorithms. Furthermore, rigid databases struggle to adapt to changing real-world data, such as shifting travel schedules or evolving city rankings.

This research proposes a paradigm shift toward a "web of trust" built on the Semantic Web. By utilizing temporal ontologies and Intuitionistic Fuzzy Sets (IFS), the authors developed a system that mimics human advice-seeking behavior, providing transparent and context-aware recommendations.

The Proposed Framework

The architecture of the system is built around autonomous agents that represent individual users. These agents interact within a decentralized network, sharing knowledge and building trust based on the quality of past interactions.

System architecture showing agents with personal temporal ontologies interacting via trust links with web services.

Key components of the framework include:

  • Temporal Ontologies: Using the Web Ontology Language (OWL) with a versioning system, the framework allows the knowledge base to evolve over time without breaking logical consistency.
  • Intuitionistic Fuzzy Sets (IFS): Unlike binary or standard fuzzy logic, IFS explicitly models membership (liking), non-membership (disliking), and uncertainty (hesitation). This allows agents to express nuanced opinions, such as being unsure about a specific recommendation.
  • Web of Trust: Agents maintain "Trust Values" for their peers. When seeking a recommendation, an agent queries its acquaintance set and weights the feedback based on established trust. This trust is dynamic; it increases when an agent provides a high-quality recommendation and decreases otherwise.

Evaluation and Results

The system was evaluated using a tourism case study involving the Places Rated Almanac, which contains ratings for 329 U.S. cities. To establish a baseline, the authors collected manual recommendations from 25 individuals to see how humans would rank these destinations for a specific demographic (tourists from India).

Comparison of manual recommendations versus system iterations across different preference lists.

As shown in the experimental results, the system-generated recommendations closely mirrored human performance. Over several iterations, the "Web of Trust" successfully filtered out low-quality suggestions, ensuring the final lists were both manageable and relevant. Because the recommendations are tied to specific trustworthy agents and clear ontological attributes, the process offers significantly higher transparency than standard matrix-factorization approaches.

Key Insights and Future Work

One of the most significant findings is that trust serves as an effective proxy for similarity. In this model, mathematical similarity is naturally absorbed into trust; agents with similar tastes will consistently provide good advice to one another, strengthening their bond. Additionally, the use of IFS allows the system to represent human-like hesitation, a feature missing in standard 5-star rating systems.

Future research will focus on the "cold start" problem—how to initialize trust for new agents—and the computational scalability of managing thousands of agents performing complex fuzzy intersections in real-time. Further testing on more rapidly evolving datasets is also required to fully stress-test the temporal ontology versioning system.

Summary updated

BibTeX
@inproceedings{bedi2007trust,
  author = {Punam Bedi and Harmeet Kaur and Sudeep Marwaha},
  title = {{Trust Based Recommender System for Semantic Web}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure