Thread
In poster session 6 in #NeurIPS2022, I will present our paper collaborated with @CMU x @GoogleAI x @UCLA.
Can we infer a user node’s preference without using any user labels? Can we instead exploit product nodes’ abundant labels given in their publicly available content?
...🧵
Can we infer a user node’s preference without using any user labels? Can we instead exploit product nodes’ abundant labels given in their publicly available content?
...🧵
1/n. Various industrial ecosystems can be presented as heterogeneous graphs, and various Heterogeneous Graph Neural Networks (HGNNs) have been proposed to summarize heterogeneity into node embeddings for better inference.
2/n. Unfortunately, there are severe label imbalance issues on HGNNs. Publicly available content nodes are abundantly labeled (e.g., text, images), whereas labels for user nodes may not be available due to privacy restrictions.
3/n. To solve this issue, transferring knowledge from external heterogeneous graphs has previously been proposed (a.k.a graph2graph transfer learning). However, most real-world heterogeneous graphs are highly proprietary; hard to get access for transfer learning.
4/n. We instead focus on another source domain to extract knowledge for zero-labeled node types: the other label-abundant node types within the same heterogeneous graph!
5/n. To solve this newly defined transfer learning problem for HGNNs, we first dissect the HGNN architecture. HGNNs provide distinct feature extractors for each node type to deal with multi-modality. 
6/n. We then look into the relations between distinct feature extractors in the same HGNN and prove we can compute the mapping matrix between source and target embedding space by exploiting their relations.
7/n. We propose Knowledge Transfer Networks (KTN) that learn the theoretically proved mapping matrix by adding an additional regularization loss to the HGNN performance loss. How simple it is! 😘
8/n. KTN outperforms state-of-the-art baselines by up to 73% across 18 transfer learning tasks. More importantly, our model can be applied to almost any HGNN models improving the performance of 6 different HGNN models by up to 960% for inference on zero-labeled node types.
n/n, n=9. Please come to our poster session and discuss how we can practically learn embeddings on heterogeneous graphs.
All credit to my amazing collaborators, @rsalakhu @phanein @JohnPalowitch @acbuller👏👏👏
All credit to my amazing collaborators, @rsalakhu @phanein @JohnPalowitch @acbuller👏👏👏
