[Paper Review] Using content features to enhance performance of user-based collaborative filtering performance of user-based collaborative filtering

Niloofar Rastin, Mansoor Zolghadri Jahromi|arXiv (Cornell University)|Feb 10, 2014

Recommender Systems and Techniques10 references3 citations

TL;DR

This paper proposes a hybrid user-based collaborative filtering method that enhances similarity measurement by incorporating content features of items, weighting ratings based on item feature similarity. Evaluated on MovieLens data, the approach improves recommendation accuracy over standard user-based CF and other baseline methods by leveraging contextual item attributes during user similarity computation.

ABSTRACT

Content-based and collaborative filtering methods are the most successful solutions in recommender systems. Content based method is based on items attributes. This method checks the features of users favourite items and then proposes the items which have the most similar characteristics with those items. Collaborative filtering method is based on the determination of similar items or similar users, which are called item-based and user-based collaborative filtering, respectively.In this paper we propose a hybrid method that integrates collaborative filtering and content-based methods. The proposed method can be viewed as user-based Collaborative filtering technique. However to find users with similar taste with active user, we used content features of the item under investigation to put more emphasis on users rating for similar items. In other words two users are similar if their ratings are similar on items that have similar context. This is achieved by assigning a weight to each rating when calculating the similarity of two users.We used movielens data set to access the performance of the proposed method in comparison with basic user-based collaborative filtering and other popular methods.

Motivation & Objective

• To address the limitations of traditional user-based collaborative filtering, which treats all ratings equally regardless of item context.
• To improve recommendation accuracy by incorporating item content features into user similarity calculations.
• To develop a hybrid approach that combines the strengths of content-based filtering and user-based collaborative filtering.
• To evaluate the proposed method's performance against standard user-based CF and other established methods.

Proposed method

• The method modifies user similarity computation by weighting ratings based on the content similarity of items being rated.
• For each pair of users, similarity is calculated using a weighted cosine similarity measure, where weights are derived from the feature similarity of items they rated.
• Content features of items are used to compute a similarity score between items, which is then applied as a weight to each rating in the similarity calculation.
• The weighting scheme ensures that ratings on items with similar features contribute more to user similarity, reflecting contextual relevance.
• The approach is implemented as an extension of standard user-based collaborative filtering, preserving its core structure while enhancing the similarity metric.
• The method is evaluated using the MovieLens dataset, comparing performance against baseline user-based CF and other popular recommendation algorithms.

Experimental results

Research questions

• RQ1 How can content features of items be effectively integrated into user-based collaborative filtering to improve recommendation accuracy?
• RQ2 What impact does weighting ratings based on item feature similarity have on user similarity computation?
• RQ3 How does the proposed hybrid method compare in performance to standard user-based collaborative filtering and other baseline methods?
• RQ4 Can the integration of content-based features mitigate the sparsity problem in user-based collaborative filtering?

Key findings

• The proposed method significantly improves recommendation accuracy compared to standard user-based collaborative filtering on the MovieLens dataset.
• Incorporating content-based weighting in user similarity computation leads to more relevant and personalized recommendations.
• The hybrid approach outperforms baseline methods, including traditional user-based CF and other state-of-the-art techniques, as measured by standard evaluation metrics.
• The use of item content features enhances the robustness of user similarity, especially in sparse rating environments.
• The results demonstrate that context-aware rating weighting leads to better prediction performance than uniform rating weighting.
• The method achieves a measurable improvement in performance, with specific quantitative gains reported in the journal publication (e.g., higher precision and recall values).