Inferring User's Preferences using Ontologies

Vincent Schickel-Zuber, Boi FaltingsAAAI 2006

Introduction

Recommender systems typically rely on Collaborative Filtering (CF) or explicit preference-based models. However, both approaches face significant hurdles. CF is plagued by the "cold-start" problem, where the system lacks sufficient data to make accurate predictions for new users or items. Preference-based systems, while more direct, often impose a high cognitive load, requiring users to navigate exhaustive surveys. This paper proposes a middle ground: the HAPPL model, which leverages Ontology Structure based Similarity (OSS) to infer user preferences from a minimal set of ratings.

Methodology: Ontology Structure based Similarity (OSS)

The core of the proposed approach is treating the item domain as an ontology—a directed acyclic graph where nodes represent concepts and edges denote "is-a" relationships. The authors introduce the A-Priori Score (APS) to quantify the specificity of a concept based on its position in the hierarchy: $APS(c) = 1 / (n_c + 2)$, where $n_c$ is the number of descendants.

When a user provides a rating, the score is propagated through the ontology using three primary mechanisms:

  1. Upward Generalization: Moving from a specific item to a broader category.
  2. Downward Specialization: Estimating preferences for specific sub-items based on general category ratings.
  3. Lowest Common Ancestor (LCA): Determining similarity between distinct items by traversing to their shared parent node.
Ontology score propagation: (a) Upward generalization, (b) Downward specialization, and (c) LCA-based similarity.

This hierarchical approach assumes that unknown features do not contribute to a score, preventing the model from overestimating a user's interest in broad categories.

Experimental Evaluation

The authors validated OSS and HAPPL using WordNet (for semantic similarity) and a custom-built movie ontology based on MovieLens data.

Key Results

The OSS metric demonstrated a 0.908 correlation with human ratings on WordNet, significantly outperforming traditional measures like Resnik (0.793) and Jiang & Conrath (0.859).

In the recommendation task, the HAPPL model excelled in "cold-start" scenarios. As shown in the performance analysis, HAPPL maintains a lower Mean Absolute Error (MAE) compared to standard Collaborative Filtering when the number of user ratings is very low (e.g., under 20).

Comparison of Mean Absolute Error (MAE) across different models. HAPPL significantly outperforms CF in cold-start scenarios (low number of ratings).

Furthermore, HAPPL achieved a novelty value of over 33%, consistently recommending non-obvious items regardless of the user's profile size. The highest overall accuracy was achieved by a hybrid model combining HAPPL with popularity-based metrics.

Discussion and Future Work

The study highlights that the structure of knowledge—how we categorize items—is as valuable as the data itself. By using the number of descendants as a proxy for specificity, the model makes more conservative and accurate inferences.

However, current limitations include the reliance on simple "is-a" relationships and the manual effort required to construct ontologies. Future research could explore multi-typed relationships (e.g., "directed-by") and automated ontology generation from metadata to enhance scalability and depth.

Summary updated

BibTeX
@inproceedings{schickelzuber2006inferring,
  author = {Vincent Schickel-Zuber and Boi Faltings},
  title = {{Inferring User's Preferences using Ontologies}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure