The paper argues that conventional measures of recommender system (RS) performance, such as RMSE and MAE, which focus on predicting exact user ratings, are suboptimal proxies for the more fundamental goal of accurately predicting user preferences. The authors propose rank-preference consistency as a more appropriate metric, which simply counts the number of prediction pairs that are inconsistent with the user's expressed product preferences.
The paper provides background on two consistency-based RS methods - unit-consistent (UC) and shift-consistent (SC) - which provably satisfy the consensus-order property, ensuring that the RS can never recommend a product that is less preferred by all users. The authors also discuss SVD-based RS methods and the GLocalK AI-based approach.
Experimental results on the MovieLens-1M and Douban-Monti datasets show that UC, SC, and GLocalK perform comparably and produce significantly fewer discordant prediction pairs (i.e., higher rank-preference consistency) than the SVD-based methods, even though the SVD variants are optimized for unitary-invariant measures like RMSE. This suggests that unitary invariance is not a fundamental property of the RS problem, and that conventional measures of performance are not suitable for evaluating RS methods.
إلى لغة أخرى
من محتوى المصدر
arxiv.org
الرؤى الأساسية المستخلصة من
by Tung Nguyen,... في arxiv.org 04-29-2024
https://arxiv.org/pdf/2404.17097.pdfاستفسارات أعمق