A Universal Framework for Offline Serendipity Evaluation in Recommender Systems via Large Language Models

Serendipity in recommender systems (RSs) has attracted increasing attention as a concept that enhances user satisfaction by presenting unexpected and useful items. However, evaluating serendipitous performance remains challenging because its ground truth is generally unobservable. The existing offline metrics often depend on ambiguous definitions or are tailored to specific datasets and RSs, thereby limiting their generalizability. To address this issue, we propose a universally applicable evaluation framework that leverages large language models (LLMs) known for their extensive knowledge and reasoning capabilities, as evaluators. First, to improve the evaluation performance of the proposed framework, we assessed the serendipity prediction accuracy of LLMs using four different prompt strategies on a dataset containing user-annotated serendipitous ground truth and found that the chain-of-thought prompt achieved the highest accuracy. Next, we re-evaluated the serendipitous performance of both serendipity-oriented and general RSs using the proposed framework on three commonly used real-world datasets, without the ground truth. The results indicated that there was no serendipity-oriented RS that consistently outperformed across all datasets, and even a general RS sometimes achieved higher performance than the serendipity-oriented RS.

Can Large Language Models Assess Serendipity in Recommender Systems?

Serendipity-oriented recommender systems aim to counteract over-specialization in user preferences. However, evaluating a user's serendipitous response towards a recommended item can be challenging because of its emotional nature. In this study, we address this issue by leveraging the rich knowledge of large language models (LLMs), which can perform a variety of tasks. First, this study explored the alignment between serendipitous evaluations made by LLMs and those made by humans. In this investigation, a binary classification task was given to the LLMs to predict whether a user would find the recommended item serendipitously. The predictive performances of three LLMs on a benchmark dataset in which humans assigned the ground truth of serendipitous items were measured. The experimental findings reveal that LLM-based assessment methods did not have a very high agreement rate with human assessments. However, they performed as well as or better than the baseline methods. Further validation results indicate that the number of user rating histories provided to LLM prompts should be carefully chosen to avoid both insufficient and excessive inputs and that the output of LLMs that show high classification performance is difficult to interpret.

Hierarchical Matrix Factorization for Interpretable Collaborative Filtering

Matrix factorization (MF) is a simple collaborative filtering technique that achieves superior recommendation accuracy by decomposing the user-item rating matrix into user and item latent matrices. This approach relies on learning from user-item interactions, which may not effectively capture the underlying shared dependencies between users or items. Therefore, there is scope to explicitly capture shared dependencies to further improve recommendation accuracy and the interpretability of learning results by summarizing user-item interactions. Based on these insights, we propose "Hierarchical Matrix Factorization" (HMF), which incorporates clustering concepts to capture the hierarchy, where leaf nodes and other nodes correspond to users/items and clusters, respectively. Central to our approach, called hierarchical embeddings, is the additional decomposition of the user and item latent matrices (embeddings) into probabilistic connection matrices, which link the hierarchy, and a root cluster latent matrix. Thus, each node is represented by the weighted average of the embeddings of its parent clusters. The embeddings are differentiable, allowing simultaneous learning of interactions and clustering using a single gradient descent method. Furthermore, the obtained cluster-specific interactions naturally summarize user-item interactions and provide interpretability. Experimental results on rating and ranking predictions demonstrated the competitiveness of HMF over vanilla and hierarchical MF methods, especially its robustness in sparse interactions. Additionally, it was confirmed that the clustering integration of HMF has the potential for faster learning convergence and mitigation of overfitting compared to MF, and also provides interpretability through a cluster-centered case study.