How Powerful is Graph Convolution for Recommendation? (CIKM’21)
上篇文章认为LightGCN的缺点是over-smoothing,这篇文章认为LightGCN的关键优点是smoothness。文章提出一个a unified graph convolution- based framework for CF,并将现有的CF方法归纳统一到这个框架中(including the neighborhood- based methods, low-rank matrix factorization, linear auto-encoders, and LightGCN, corresponding to different low-pass filters. ),进而提出GF-CF模型,(Graph Filter based Collaborative Filtering)。效果提升70%,训练时间只需要65.8秒。
untrained LightGCN的威力
LightGCN效果好是因为smoothness,smoothness可以定义为有连接的两个点的表征的相似度:
S 2 ( x ) = ∑ i , j A i , j ∥ x i − x j ∥ 2 2 S_{2}(x)=\sum_{i, j} A_{i, j}\left\|x_{i}-x_{j}\right\|_{2}^{2} S2(x)=i,j∑Ai,j∥xi−xj∥22
当LightGCN的模型维度达到一定规模时,即使是随机初始化的未经过训练,也可以超过64维的精心训练的LightGCN:
作者认为这一定是“**linear filter playing the essential role **”。
但是这也是有一定条件的:模型维度和数据集稠密度正相关,这也是为什么在更稀疏的Amazon-book上效果好。
作者证明出,当 K K K 层的LightGCN模型维度趋近于无穷大时,模型的为用户 u u u 的推荐等价于:
s u = ∑ k = 0 K − 1 β k r ~ u ( R ~ T R ~ ) k \boldsymbol{s}_{u}=\sum_{k=0}^{K-1} \beta_{k} \tilde{\boldsymbol{r}}_{u}\left(\tilde{\boldsymbol{R}}^{T} \tilde{\boldsymbol{R}}\right)^{k} su=k=0∑K−1βkr~u