openhgnn.models.SeHGNN

class SeHGNN(args)

This is a model SimpleHGN from Simple and Efficient Heterogeneous Graph Neural Network This model is a metapath-based model. It put the neighbor aggregation in the preprocessing step, and using the single-layer structure and long metapaths. It performed over the state-of-the-arts on both accuracy and training speed.

the neighbor aggregation

\[\mathrm{X}^{P} = \hat{A}_{c,c_{1}}\hat{A}_{c_{1},c_{2}}...\hat{A}_{c_{l-1},c_{l}} \mathrm{X}^{c_{l}}\]

feature projection

\[{\mathrm{H}^{'}}^{P} = MLP_{P}(\mathrm{X}^{P})\]

semantic fusion (transformer):

\[\begin{split}q^{\mathcal{P}_{i}}=W_{Q} h^{\prime \mathcal{P}_{i}}, k^{\mathcal{P}_{i}}=W_{K} h^{\prime \mathcal{P}_{i}}, v^{\mathcal{P}_{i}}=W_{V} h^{\prime \mathcal{P}_{i}}, \mathcal{P}_{i} \in \Phi_{X} \\\end{split}\]

\[\alpha_{\left(\mathcal{P}_{i}, \mathcal{P}_{j}\right)}=\frac{\exp \left(q^{\mathcal{P}_{i}} \cdot k^{{\mathcal{P}_{j}}^{T}}\right)}{\sum_{\mathcal{P}_{t} \in \Phi_{X}} \exp \left(q^{\mathcal{P}_{i}} \cdot k^{{\mathcal{P}_{t}}^{T}}\right)}\]

\[h^{\mathcal{P}_{i}}=\beta \sum_{\mathcal{P}_{j} \in \Phi_{X}} \alpha_{\left(\mathcal{P}_{i}, \mathcal{P}_{j}\right)} v^{\mathcal{P}_{j}}+h^{\prime \mathcal{P}_{i}}\]