Mixed Collaborative and Content-Based Filtering with User-Contributed Semantic Features

Matthew Garden, Gregory DudekAAAI 2006

Introduction

Traditional recommender systems typically rely on either Collaborative Filtering (CF) or Content-Based Filtering (CBF). While CF is popular for its simplicity, it suffers from the "cold start" problem and an inherent information gap—relying solely on sparse star ratings to model complex human preferences. This paper introduces the Recommendz system, which bridges this gap by incorporating user-contributed "semantic features" (tags) to create richer profiles of both users and items.

The Recommendz Framework

Unlike systems with rigid taxonomies, Recommendz allows users to define their own descriptive features (e.g., "gritty," "slapstick"). When rating an item, users provide two dimensions of feedback for each feature:

  1. Quantity ($q$): The perceived amount of the feature in the item (0–10).
  2. Opinion ($o$): The user's preference for that feature in that specific context (-5 to +5).

To handle data sparsity, the system employs semantic binning, categorizing feature quantities into "low," "medium," and "high" based on user-specific statistical distributions.

Feature-Based Filtering (FB)

The core of the system is a probabilistic Feature-Based (FB) model. It calculates a user's expected rating for an item by averaging their historical preferences for the specific features associated with that item. This approach allows the system to make recommendations based on the underlying attributes that drive user satisfaction, rather than just raw rating correlations.

Personalized Algorithm Selection

A key insight of this research is that no single recommendation algorithm is optimal for all users. The authors implemented a "Combined" hybrid approach that personalizes the choice of algorithm. By using a validation set, the system identifies whether FB, CF, or another variant performs best for a specific individual.

Distribution of optimal recommendation methods across the user base.

As shown in the data, the Feature-Based method was the superior choice for over 60% of users, highlighting the value of semantic features over traditional collaborative methods for the majority of the population.

Experimental Results

The system was evaluated using a dataset of 885 users and 1,641 movies. Performance was measured using Normalized Mean Absolute Error (NMAE) and Spearman rank correlation ($\rho$).

Comparative performance of recommendation algorithms across NMAE, Spearman rho, and Coverage.

The results demonstrate that the Combined hybrid approach achieves the highest accuracy (lowest NMAE) and the best rank-ordering (highest Spearman $\rho$). While the FB method significantly outperforms pure CF in accuracy, it does so at a slight cost to coverage (87% vs 94%), as it requires overlapping tags between users and items to generate a prediction.

Conclusion

By leveraging user-contributed semantic features and personalizing the recommendation algorithm itself, the Recommendz system offers a significant improvement over standard collaborative filtering. Future work will focus on improving coverage through synonym detection and exploring non-linear models to better capture complex user preferences.

Summary updated

BibTeX
@inproceedings{garden2006mixed,
  author = {Matthew Garden and Gregory Dudek},
  title = {{Mixed Collaborative and Content-Based Filtering with User-Contributed Semantic Features}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure