Perturbation-Recovery Method for Recommendation

17 Nov 2022  ·  Jeongwhan Choi, Seoyoung Hong, Noseong Park, Sung-Bae Cho ·

Collaborative filtering is one of the most influential recommender system types. Various methods have been proposed for collaborative filtering, ranging from matrix factorization to graph convolutional methods. Being inspired by recent successes of GF-CF and diffusion models, we present a novel concept of blurring-sharpening process model (BSPM). Diffusion models and BSPMs share the same processing philosophy in that new information is discovered (e.g., a new image is generated in the case of diffusion models) while original information is first perturbed and then recovered to its original form. However, diffusion models and our BSPMs deal with different types of information, and their optimal perturbation and recovery processes have a fundamental discrepancy. Therefore, our BSPMs have different forms from diffusion models. In addition, our concept not only theoretically subsumes many existing collaborative filtering models but also outperforms them in terms of Recall and NDCG in the three benchmark datasets, Gowalla, Yelp2018, and Amazon-book. Our model marks the best accuracy in them. In addition, the processing time of our method is one of the shortest cases ever in collaborative filtering. Our proposed concept has much potential in the future to be enhanced by designing better blurring (i.e., perturbation) and sharpening (i.e., recovery) processes than what we use in this paper.

PDF Abstract

Datasets


Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Recommendation Systems Amazon-Book BSPM-EM Recall@20 0.0733 # 1
nDCG@20 0.0609 # 2
Recommendation Systems Amazon-Book BSPM-LM Recall@20 0.0733 # 1
nDCG@20 0.0610 # 1
Collaborative Filtering Gowalla BSPM-LM Recall@20 0.1878 # 2
Recommendation Systems Gowalla BSPM-EM Recall@20 0.1910 # 1
nDCG@20 0.1581 # 1
Recommendation Systems Gowalla BSPM-LM Recall@20 0.1878 # 2
nDCG@20 0.1548 # 5
Collaborative Filtering Gowalla BSPM-EM Recall@20 0.1910 # 1
Recommendation Systems Yelp2018 BSPM-LM Recall@20 0.0713 # 2
NDCG@20 0.0584 # 2
Recommendation Systems Yelp2018 BSPM-EM Recall@20 0.0719 # 1
NDCG@20 0.0591 # 1

Methods


No methods listed for this paper. Add relevant methods here