Reproduce “PMD: An Optimal Transportation-based User Distance for Recommender Systems” on OpenResearch

About this paper

Yitong Meng, Xinyan Dai, Xiao Yan, James Cheng, Weiwen Liu, Benben Liao, Jun Guo, Guangyong Chen

Preference Mover's Distance (PMD) is presented as a novel user distance metric for recommender systems, designed to address data sparsity by framing user similarity as an Earth Mover's Distance problem. This approach consistently achieved superior recommendation accuracy (lower MAE) on MovieLens datasets compared to traditional and advanced methods, particularly in highly sparse conditions.

Key insights

• By casting user similarity as an EMD problem, PMD ensures the resulting user distance is a metric, providing desirable properties like the triangle inequality for more consistent and reliable similarity values.
• PMD effectively overcomes the limitation of requiring co-rated items, enabling robust user similarity computation even when users have rated completely different item sets.
• The EMD framework provides a natural and principled normalization strategy for user similarity, which prevents the misleading scaling issues observed in some heuristic methods.

Summary by alphaXiv.