RUCAIBox / NCL

[WWW'22] Official PyTorch implementation for "Improving Graph Collaborative Filtering with Neighborhood-enriched Contrastive Learning".
119 stars 20 forks source link

参数设置请教 #45

Closed shaohuacongwen closed 10 months ago

shaohuacongwen commented 1 year ago

我们注意到您论文中sgl模型与sgl原文结果差距较大,想向您请教一下您调的sgl参数

hyp1231 commented 1 year ago

SGL 原文的数据处理是:we use the same 10-core setting for Yelp2018 and Amazon-Book ... split the interactions into training, validation, and testing set with a ratio of 7:1:2 ...

而 NCL 的数据处理主要使用 RecBole 框架自带的数据划分模块,且数据处理为:For Yelp and Amazon Books datasets, we filter out users and items with fewer than 15 interactions to ensure data quality. ... For each dataset, we randomly select 80% of interactions as training data and 10% of interactions as validation data. The remaining 10% interactions are used for performance comparison.

所以两篇论文的实验设置并不相同,直接比较两篇论文的指标数值并无实际意义。

shaohuacongwen commented 1 year ago

我们希望复现您文章中sgl的结果,以做进一步分析,想请教下具体参数和模型设置

hyp1231 commented 1 year ago

我们跑 baseline 时使用的 SGL 模型文件是 sgl.py,应该比较好迁移到当前代码框架下。

在跑 baseline 时,我们汇报的是超参调优后的结果,但是当时好像没有记录最优参数和调参范围,推荐您尝试使用 RecBole 自带的超参调优工具。具体可参考 #40 。