[Paper Review] Two-Way Latent Grouping Model for User Preference Prediction

Eerika Savia, Kai Puolamäki+2 more|arXiv (Cornell University)|Jul 4, 2012

Recommender Systems and Techniques7 references3 citations

TL;DR

This paper proposes a Two-Way Latent Grouping Model that jointly models latent user and document groups to improve user preference prediction, especially for new documents with few ratings. Using Gibbs sampling, the model outperforms the User Rating Profile model by leveraging document-level grouping, significantly enhancing generalization and prediction accuracy in cold-start scenarios for new documents.

ABSTRACT

We introduce a novel latent grouping model for predicting the relevance of a new document to a user. The model assumes a latent group structure for both users and documents. We compared the model against a state-of-the-art method, the User Rating Profile model, where only users have a latent group structure. We estimate both models by Gibbs sampling. The new method predicts relevance more accurately for new documents that have few known ratings. The reason is that generalization over documents then becomes necessary and hence the twoway grouping is profitable.

Motivation & Objective

• To improve user preference prediction accuracy for new documents with sparse rating data.
• To address the cold-start problem in recommendation systems where new items have few known ratings.
• To model both user and document latent group structures simultaneously, rather than only user groups.
• To enhance generalization by incorporating document-level clustering alongside user-level clustering.
• To evaluate the model's performance against a state-of-the-art baseline (User Rating Profile) in relevance prediction.

Proposed method

• The model assumes a latent group structure for both users and documents, enabling joint clustering of both entities.
• It employs a generative probabilistic framework where user and document group memberships are inferred from observed ratings.
• Gibbs sampling is used to estimate the posterior distribution over latent group assignments for users and documents.
• The model learns user preferences by considering both their latent group and the latent group of the document being rated.
• The likelihood of a rating is modeled conditionally on the joint user and document group memberships.
• The two-way grouping allows for better generalization when predicting relevance for new documents with limited ratings.

Experimental results

Research questions

• RQ1 Can jointly modeling latent user and document groups improve prediction accuracy for new documents with few ratings?
• RQ2 How does the inclusion of document-level grouping affect performance compared to models that only group users?
• RQ3 Does the two-way latent grouping model generalize better in cold-start scenarios for new documents?
• RQ4 What is the relative contribution of user and document group structures to prediction performance?
• RQ5 How does the model's performance compare to the User Rating Profile model, which only models user groups?

Key findings

• The Two-Way Latent Grouping Model significantly improves prediction accuracy for new documents with few known ratings compared to the User Rating Profile model.
• The model's performance gain is most pronounced in cold-start scenarios, where document-level generalization is essential.
• Joint modeling of user and document groups enables better generalization, as the model leverages patterns across both dimensions.
• The use of Gibbs sampling enables effective inference over the latent group structures for both users and documents.
• The results demonstrate that incorporating document group structure is beneficial when predicting relevance for items with sparse rating data.
• The model achieves superior performance in relevance prediction, particularly when the number of known ratings per document is low.