Self supervised Nodes Hyperedges Embedding for Heterogeneous Information Network Learning

reviews

Summary:

This work studies self-supervised learning problem on heterogeneous graphs. It converts heterogeneous graph into hypergraph through metapaths, then adopts HGCN to obtain representation. It also proposes a dual self-supervised structure, to minimize the KL divergence of two distributions from different model predictions from different views.

  1. The idea to convert heterogeneous graph to hypergraph via metapaths to better capture higher order information is intereseting.
  2. The experiments show promising results, with adequate ablation study.

  3. This presentation is clear.

  4. The proposed method for node classification can hardly be considered "self-supervised learning" if it contains \(L_{CE}\) in the training loss term \(L=L_{CE}+\alpha L_{dnn} +\beta L_{hgcn}\). The typical evaluation of self-supervised node classification should be an unsupervised learning representation (trained by self-supervised learning loss) followed by a linear supervised classification layer, as in most graph self-supervised learning papers, e.g., [1].

  5. The graph datasets used in experiments are relatively small. To elaborate the scalability and performance on larger graphs, it will be better to discuss the time/space complexity and include experiments on graphs of larger size, e.g., BGS(916k edges) and AM(5.9M edges) in [2].
  6. The two different views used in dual self-supervised structure are a bit ad-hoc. Better insights and motivation should be provided here, and more advanced graph view generation methods should also be considered and discussed.

[1] Self-Supervised Learning of Graph Neural Networks: A Unified Review, Yaochen Xie, Zhao Xu, Jingtun Zhang, Zhengyang Wang, Shuiwang Ji, TPAMI 2021 [2] Modeling Relational Data with Graph Convolutional Networks, Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov & Max Welling, ESWC 2018

It is better to include variance in Table 3 and 4, obtained from repetitions of experiments.

typos: Section 5.2 Beaselines -> Baselines Table 5 hpyergraph-> hypergraph

This paper gives an interesting idea to take heterogeneous graphs as hypergraphs to better capture higher order information.

However, the experiment evaluation has some flaws, including the evaluation method and the scalability. And it is encouraged to provide better motivation and insights about the design of dual self-supervised structure, and to discuss more methods for the generation of the different views (which are simple F and Z in this paper).