[Paper Review] Using Contextual Information as Virtual Items on Top-N Recommender Systems

Marcos Aurélio Domingues, Alí­pio Jorge+1 more|arXiv (Cornell University)|Jan 1, 2009

Recommender Systems and Techniques14 references16 citations

TL;DR

This paper proposes DaVI (Dimensions as Virtual Items), a method to enhance Top-N recommender systems by treating contextual attributes—such as time, location, and music genre—as virtual items within existing collaborative filtering and association rule algorithms. By integrating contextual data directly into the recommendation model without modifying the core algorithm, DaVI improves prediction accuracy, especially when rich contextual dimensions (e.g., band or genre) are available, with gains up to 34% in F1 score over traditional user-item models.

ABSTRACT

Traditionally, recommender systems for the Web deal with applications that have two dimensions, users and items. Based on access logs that relate these dimensions, a recommendation model can be built and used to identify a set of N items that will be of interest to a certain user. In this paper we propose a method to complement the information in the access logs with contextual information without changing the recommendation algorithm. The method consists in representing context as virtual items. We empirically test this method with two top-N recommender systems, an item-based collaborative filtering technique and association rules, on three data sets. The results show that our method is able to take advantage of the context (new dimensions) when it is informative.

Motivation & Objective

• To improve the accuracy of Top-N recommender systems by incorporating contextual information without altering existing recommendation algorithms.
• To investigate whether contextual dimensions such as time, location, and music genre can be effectively used to enhance recommendation performance.
• To evaluate the effectiveness of representing context as virtual items in both item-based collaborative filtering and association rule-based recommender systems.
• To compare DaVI with existing contextual recommendation approaches, such as Combined Reduction, in multi-dimensional context scenarios.

Proposed method

• Represent each contextual attribute (e.g., hour, day, genre) as a virtual item in the user-item interaction matrix.
• Integrate contextual dimensions into the recommendation model by treating them as additional items, preserving the structure of existing algorithms.
• Apply the DaVI method to two recommendation techniques: item-based collaborative filtering (using cosine similarity) and association rule mining (with minimum support and confidence thresholds).
• Use forward selection and full combination strategies to identify optimal sets of contextual dimensions for recommendation.
• Compare DaVI with baseline models (user-item only) and a modified Combined Reduction approach that segments sessions by context.
• Evaluate performance using the F1 measure on three real-world datasets: Listener, Playlist, and Entree Chicago restaurant data.

Experimental results

Research questions

• RQ1 Can treating contextual attributes as virtual items improve the predictive accuracy of Top-N recommender systems?
• RQ2 How does the performance of DaVI vary across different types of contextual dimensions (e.g., inferred from logs vs. extracted from CMS)?
• RQ3 Does combining multiple contextual dimensions through DaVI outperform using individual dimensions or baseline models?
• RQ4 How does DaVI compare to the Combined Reduction approach in terms of F1 score and scalability?
• RQ5 Which contextual dimensions yield the highest gains in recommendation accuracy, and under what conditions?

Key findings

• DaVI improved the F1 score by up to 34% in the Listener dataset using the 'band' dimension with item-based collaborative filtering.
• In the Playlist dataset, DaVI achieved a 24% F1 gain using the 'band' dimension, demonstrating strong performance with rich contextual data.
• The Entree dataset showed only a 5% F1 gain with the 'intention' dimension, indicating that less informative contexts yield smaller improvements.
• Association rule-based models outperformed item-based CF in the Entree dataset, achieving a maximum F1 of 0.341 with the 'intention' dimension.
• DaVI with forward selection and all dimensions performed competitively, with the best results often matching or exceeding the 'best context' individual dimension approach.
• The Combined Reduction baseline outperformed DaVI in some cases, particularly with association rules, suggesting trade-offs in model complexity and context integration.