[Paper Review] MeLU: Meta-Learned User Preference Estimator for Cold-Start Recommendation

Hoyeop Lee, Jinbae Im+3 more|arXiv (Cornell University)|Jul 31, 2019

Recommender Systems and Techniques31 references53 citations

TL;DR

MeLU uses optimization-based meta-learning (MAML) to rapidly personalize recommendations for new users with very few interactions, and introduces an evidence-candidate selection strategy to improve initial recommendations in cold-start scenarios.

ABSTRACT

This paper proposes a recommender system to alleviate the cold-start problem that can estimate user preferences based on only a small number of items. To identify a user's preference in the cold state, existing recommender systems, such as Netflix, initially provide items to a user; we call those items evidence candidates. Recommendations are then made based on the items selected by the user. Previous recommendation studies have two limitations: (1) the users who consumed a few items have poor recommendations and (2) inadequate evidence candidates are used to identify user preferences. We propose a meta-learning-based recommender system called MeLU to overcome these two limitations. From meta-learning, which can rapidly adopt new task with a few examples, MeLU can estimate new user's preferences with a few consumed items. In addition, we provide an evidence candidate selection strategy that determines distinguishing items for customized preference estimation. We validate MeLU with two benchmark datasets, and the proposed model reduces at least 5.92% mean absolute error than two comparative models on the datasets. We also conduct a user study experiment to verify the evidence selection strategy.

Motivation & Objective

• Address user cold-start in recommender systems where initial interactions are scarce.
• Leverage meta-learning to produce personalized preference estimators per user.
• Propose an evidence candidate selection strategy to improve early recommendation quality.
• Validate performance on benchmark datasets and via a user study to assess evidence selection.

Proposed method

• Define a user preference estimator that embeds user and item contents and passes them through a multi-layer neural decision-making network.
• Adopt an optimization-based meta-learning (MAML) framework to rapidly adapt model parameters for each new user using a support set (their item history).
• Do not update user/item embeddings during local updates to maintain stability; only update the decision layers and output layer.
• Train with a two-level update: local updates using each user’s support set, followed by a global update across tasks using query sets.
• Introduce an evidence-candidate selection strategy that scores items by combining gradient-based distinction (Frobenius norm of personalization gradient) and item popularity to form top-k evidence candidates.
• Evaluate MeLU on MovieLens 1M and BookCrossing datasets across four scenarios (existing/new items and existing/new users) using MAE and nDCG as metrics; compare to PPR and Wide & Deep baselines.

Experimental results

Research questions

• RQ1 Can a MAML-based recommender rapidly personalize to a new user with only a small amount of interaction data?
• RQ2 Does selectively choosing evidence candidates improve initial recommendations for new users compared to popularity-based candidates?
• RQ3 How robust is MeLU to varying lengths of a user’s item-consumption history in cold-start settings?
• RQ4 Do the learned personalized parameters generalize across different datasets and cold-start scenarios?

Key findings

• MeLU outperforms two baseline models in three cold-start scenarios across MovieLens and BookCrossing datasets.
• The model adapts quickly, with substantial MAE improvements after a single local update; additional updates offer diminishing returns.
• Evidence-candidate selection based on personalization gradients and popularity leads to more reliable candidates and higher user satisfaction in a user study.
• MeLU maintains strong performance even with very short item-consumption histories, demonstrating robustness to history length.