[Model]add HGOT

[zhuyanhuazhuyanhua]

HGOT[ICML2025]

• paper: HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport
• Code from author: HGOT

How to run

• Clone the Openhgnn-DGL

  python main.py -m HGOT -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 & Feature Transformation
  • To handle the heterogeneity where nodes of different types possess features of varying dimensions (or exist in different feature spaces), HGOT first projects the raw features into a unified latent space.
  • Unified Dimension Projection: Type-specific linear transformation matrices are designed to map the original features of each node type into a shared vector space.
  • Formula: For the i-th node of type $\phi_i$,the projected feature $h_i$
    is calculated as:$h_i$ = $ W_{\phi_i}$ * $x_i$ + $ b_{\phi_i} $
• Heterogeneous Semantic Information Aggregation
  • To capture the rich high-order semantic information within the heterogeneous graph, HGOT employs an attention mechanism to aggregate multiple meta-path based views into a comprehensive "central view".
  • Branch View Generation: The original heterogeneous graph is decomposed into several homogeneous subgraphs (i.e., branch views) based on different meta-paths.
  • Central View Construction:
    Node Level: An attention network is utilized to compute node embeddings under each meta-path, which are then fused via weighted summation to obtain the aggregated node representation $ Z_agg $
    Edge Level: The adjacency matrices generated by each meta-path undergo a logical OR operation to produce the aggregated adjacency matrix $ A_agg $, preserving all potential connectivity information.
  • Objective: The central view integrates semantic information from various branch views and serves as the target distribution in the optimal transport process.

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

Yanhua Zhu[GAMMA LAB]

If you have any questions,

Submit an issue or email to 2023210015@bupt.cn.

More

Contirbutor

Yanhua Zhu[GAMMA LAB]

If you have any questions,

Submit an issue or email to 2023210015@bupt.cn.

Description

Checklist

Please feel free to remove inapplicable items for your PR.

• [Y] The PR title starts with [$CATEGORY] (such as [NN], [Model], [Doc], [Feature]])
• [Y] Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage
• Code is well-documented
• To the best of my knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change
• Related issue is referred in this PR
• If the PR is for a new model/paper, I've updated the example index here.

Changes