GCN現有變體彙總(應用篇)
Mix Hop(高階多跳的圖特徵)融合
文獻:
ICML_2019
MixHop: Higher-Order Graph Convolutional Architectures via Sparsified Neighborhood Mixing
AAAI_20: Multi-Range Attentive Bicomponent Graph Convolutional Network for Traffic Forecasting
2. 兩路並行,同時對節點和邊的關係建模,形成以邊爲中心的圖網絡和以節點爲中心的圖網絡
NodeNet
EdgeNet
AAAI_20: Multi-Range Attentive Bicomponent Graph Convolutional Network for Traffic Forecasting (同上)
TKDE_20: Flow Prediction in Spatio-Temporal Networks Based on Multitask Deep Learning
IJCAI_19: MR-GNN: Multi-Resolution and Dual Graph Neural Network for Predicting Structured Entity Interactions
兩路並行 將GCN的卷積結果和S-LSTM(summary)和I-LSTM(interaction)
1. weighted graph convolution
2. graph-gather layers 經過一層全連接再加起來得到全圖的全部信息(和)是表示graph-level的信息
3. 對gt做 graph-state的S-LSTM 也就是對summary graph-gate做 graph-level的LSTM
4. 對gXt和gYt進行連接,再對其做LSTM 就是interaction
5. 最後把得到的結果都concantenate起來 經過全連接 得到1*k的向量 k表示標註交集的label數。
這個工作得到的都是graph-level的結果,我們也可以拓展到node-level去
3. Multi-step Prediction: GCN+Seq2Seq
IJCAI_19: STG2Seq: Spatial-Temporal Graph to Sequence Model for Multi-step Passenger Demand Forecasting
IJCAI_19: GSTNet: Global Spatial-Temporal Network for Traffic Flow Prediction
STSGCN https://github.com/Davidham3/STSGCN AAAI_20 【Spatial-Temporal Synchronous Graph Convolutional Networks: A New Framework for Spatial-Temporal Network Data Forecasting】
It is designed for spatio-temporal network data forecasting, which captures complex localized spatial-temporal correlations and heterogeneity with a Spatial-Temporal Synchronous Graph Convolutional Network.
4. 異質GCN:Hetero-GCN
KDD_19: Heterogeneous Graph Neural Network
AAAI_20: An Attention-based Graph Neural Network for Heterogeneous Structural Learning
推薦系統裏建模異質網絡IntentGC
IntentGC: a Scalable Graph Convolution Framework Fusing Heterogeneous Information for Recommendation
vector-wise/bit-wise
5. MaskGCN:
IJCAI_19: STAR-GCN: Stacked and Reconstructed Graph Convolutional Networks for Recommender Systems
IJCAI_19: Masked Graph Convolutional Network
Network embedding就是通過訓練特徵表示representation來使得圖中相鄰的節點表徵儘可能小,而較遠的節點表徵儘可能大。或者使得特徵表示滿足其他的task相關的要求。
Network embedding aims to represent graph nodes in a low dimensional space where the network structure and properties are preserved.