Core Concepts
The author proposes the Lower-Left Partial AUC (LLPAUC) as an efficient optimization metric that correlates strongly with Top-K ranking metrics, addressing the limitations of existing metrics.
Abstract
The paper introduces LLPAUC as a novel optimization metric for recommendation systems. It highlights the correlation between LLPAUC and Top-K ranking metrics, providing theoretical validation and empirical evidence. The proposed loss function is compared against various baselines in clean and noisy training settings across different datasets, demonstrating superior performance.
The study emphasizes the importance of optimizing recommendation systems efficiently while aligning with key performance metrics like Recall@K and NDCG@K. By introducing LLPAUC, the authors offer a promising approach to enhance recommendation quality and robustness against noisy feedback.
Stats
LLPAUC exhibits better performance than existing metrics on Adressa, Yelp, and Amazon datasets.
LLPAUC achieves higher Recall@20 and NDCG@20 scores compared to BPR, BCE, SCE, CCL, DNS(ð, ð), Softmax_v(ð, ð), PAUCI(OPAUC), LightGCN.
The LLPAUC surrogate loss function shows consistent improvement across different backbones (MF and LightGCN).
Quotes
"The proposed LLPAUC metric exhibits a stronger correlation with Top-K ranking metrics."
"LLPAUC enhances model robustness against noise in recommender systems."