推荐tomkat0789 · 2021-03
“EDIT to add another funny "frustration" paper more directly related to ML [1].”
补充一篇有趣的“挫败”类论文,说明许多神经推荐方法打不过简单基线
多数神经推荐算法难复现,且常被简单基线超越
论文系统分析了近年顶会上提出的 18 个 top-n 推荐算法,其中只有 7 个能以合理的努力复现;这 7 个里又有 6 个常被基于最近邻或图的简单启发式方法超越,剩下的一个明显优于基线,但也没有稳定地胜过调好的非神经线性排序方法。作者借此指出当前机器学习研究实践中的若干问题,呼吁改进科学规范。
评论者用它说明神经网络推荐并不比简单、调好的传统方法强,并提醒这是 RecSys 2019 最佳论文,而相关博客文章没有提到。
“EDIT to add another funny "frustration" paper more directly related to ML [1].”
“(Best Paper Award, RecSys 2019). It is strange that the blog post fails to mention this.”
“some voices[1] argue that traditional or much simpler approaches still often do a similar job”
“neural network-based approaches didn't seem to fair too well, either: Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches (arXiv pre-print, but this is an award-winning paper).”