UltraGCN: Ultra Simplification of Graph Convolutional Networks for Recommendation (CIKM’21 华为诺亚方舟)
SOTA之一
(文章放在草稿箱忘发出来了,放了两个多月)
文章分析了LightGCN的三个缺点,并提出了一种极端简化的GCN推荐模型,几乎就是degree weighted MF,没有message passing、aggregation,用两个辅助loss把图结构信息引入MF,要把Amazon-Book数据集刷爆了,速度是LightGCN的15倍,效果提升60%,可谓大道至简,返璞归真,感觉就是MF+ItemCF的模型融合。同时也有伪标签的思想在里面。
LightGCN三大缺点
缺点一:消息聚合的边权重是违反直觉的
LightGCN的消息聚合:
e u ( l + 1 ) = 1 d u + 1 e u ( l ) + ∑ k ∈ N ( u ) 1 d u + 1 d k + 1 e k ( l ) e i ( l + 1 ) = 1 d i + 1 e i ( l ) + ∑ v ∈ N ( i ) 1 d v + 1 d i + 1 e v ( l ) \begin{aligned} &e_{u}^{(l+1)}=\frac{1}{d_{u}+1} e_{u}^{(l)}+\sum_{k \in \mathcal{N}(u)} \frac{1}{\sqrt{d_{u}+1} \sqrt{d_{k}+1}} e_{k}^{(l)} \\ &e_{i}^{(l+1)}=\frac{1}{d_{i}+1} e_{i}^{(l)}+\sum_{v \in \mathcal{N}(i)} \frac{1}{\sqrt{d_{v}+1} \sqrt{d_{i}+1}} e_{v}^{(l)} \end{aligned} eu(l+1)=du+11eu(l)+k∈N(u)∑du+1dk+11ek(l)ei(l+1)=di+11ei(l)+v∈N(i)∑dv+1di+11ev(l)
点积时:
e u ( l + 1 ) ⋅ e i ( l + 1 ) = α u i ( e u ( l ) ⋅ e i ( l ) ) + ∑ k ∈ N ( u ) α i k ( e i ( l ) ⋅ e k ( l ) ) + ∑ v ∈ N ( i ) α u v ( e u ( l ) ⋅ e v ( l ) ) + ∑ k ∈ N ( u ) ∑ v ∈ N ( i ) α k v ( e k ( l ) ⋅ e v ( l ) ) \begin{gathered} e_{u}^{(l+1)} \cdot e_{i}^{(l+1)}=\alpha_{u i}\left(e_{u}^{(l)} \cdot e_{i}^{(l)}\right)+\sum_{k \in \mathcal{N}(u)} \alpha_{i k}\left(e_{i}^{(l)} \cdot e_{k}^{(l)}\right)+ \\ \sum_{v \in \mathcal{N}(i)} \alpha_{u v}\left(e_{u}^{(l)} \cdot e_{v}^{(l)}\right)+\sum_{k \in \mathcal{N}(u)} \sum_{v \in \mathcal{N}(i)} \alpha_{k v}\left(e_{k}^{(l)} \cdot e_{v}^{(l)}\right) \end{gathered} eu(l+1)⋅ei(l+1)=αui(eu(l)⋅ei(l))+k∈N(u)∑α