HGNN_AC

HGNN_AC[WWW2021]

• paper: Heterogeneous Graph Neural Network via Attribute Completion
• Code from author: HGNN_AC

How to run

• Clone the Openhgnn-DGL

  python main.py -m HGNN_AC -t node_classification -d imdb4MAGNN -g 0

  If you do not have gpu, set -gpu -1.

  the dataset imdb4MAGNN is supported.

Performance: Node classification

• Device: CPU, Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz 2.59 GHz
• Dataset: IMDB

Node classification	Macro-F1	Micro-F1
MAGNN	58.65%	59.20%
paper	60.75%	60.98%
OpenHGNN	60.54%	60.70%

The perform of experiments are run in the setting of paper which uses SVM classification, so it is a little bit different from semi-supervised node classification. And directly running the model is using semi-supervised node classification trainerflow.

Dataset

• We process the IMDB dataset given by MAGNN. It is saved as dgl.heterograph and can be loaded by dgl.load_graphs

Description

• imdb4MAGNN

  • Number of nodes

    movie	4278
    director	2081
    actor	5257

  • Number of edges

    movie-director	4278
    movie-actor	12828

  • Types of metapaths: MDM, MAM, DMD, DMAMD, AMA, AMDMA. Please note that the M is movie, D is director, A is actor, and the edges above are all bidirectional.

[TODO]

TrainerFlow: Node classification trainer

• Graph preprocess

  • To get some information from the original graph. It divides the nodes with attributes into a feature keep list and a feature drop list. It also uses the edge information to get the nodes' adjacency matrices related to the source nodes.

• Attribute Completion with Attention Mechanism

  • HGNN-AC adopts a masked attention mechanism which means we only calculate $e_{vu}$ for nodes $u\in{N_v^+}$, where $u\in{N_v^+}$ denotes the first-order neighbors of node $v$ in set $V^+$, where $V^+$ is the set of nodes with attributes.
  • Then, softmax function is applied to get normalized weighted coefficient $a_{vu}$.
  • HGNN-AC can perform weighted aggregation of attributes for node $v$ according to weighted coefficient $a_{vu}$:
  • Specially, the attention process is extended to a multi-head attention to stabilize the learning process and reduce the high variance

• Dropping some Attributes

  • HGNN-AC drops some attributes according to the feature drop list. It uses the Attribute Completion to get the new features of the nodes and compares the new features with the original features.

• Combination with the HIN Model

  • Finally, we can apply the proposed framework to MAGNN model or some other models. For more details, you can refer to the corresponding models.

Hyper-parameters specific to the model

You can modify the parameters in openhgnn/config.ini

Description

feats_drop_rate = 0.3 # feature drop rate to get the feature drop list
attn_vec_dim = 64 # the dimesions of vector in the Attention Layer 
feats_opt = 110 # the type of nodes that needs to get the new features
loss_lambda = 0.2 # the weighted coefficient to balance the two parts.
src_node_type = 2 # the type of nodes that has the raw attributes
dropout = 0.1 # the drop rate used in Drop some Attributes
num_heads = 8 # the num of heads used in muti-head attention mechanism
HIN = MAGNN # the type of model used in Combination with the HIN Model.

More

Contirbutor

Yaoqi Liu[GAMMA LAB]

If you have any questions,

Submit an issue or email to YaoqiLiu@bupt.edu.cn.