wendongbi / Bridged-GNN

Bridged-GNN for Knowledge Bridge Learning (KBL), the paper is accepted by CIKM2023

Bridged-GNN

Bridged-GNN for Knowledge Bridge Learning (KBL), the paper is accepted by CIKM2023.

Datasets

We conduct experiments on five datasets, including four real-world datasets and a synthetic dataset. We preprocessed all datasets used in this paper, including relational and unrelational datasets, into the format of torch_geometric. And our processed datasets are available at ./dataset/xxx.dat (unzip the file named 'dataset_pyg.tar.gz' in the dataset folder first). And we provide a demo to load the dataset. The detailed information of these datasets are as follows:

• Twitter: Twitter is a social network dataset that describes the social relations among politicians (source domain) and civilians (target domain) on the Twitter platform. The node features are the GloVe embedding of user's Tweets (we do not use the user profile features of politicians used in the original paper), and the goal is to predict the political tendency of civilians. The original dataset ($\text{Twitter}\text{Graph}$) is used in $RD{intra-and-inter}$ scenarios. By removing all original edges, we also construct the $\text{Twitter}_{UD}$ dataset in un-relational data ($UD$) scenarios.

• Office31: The Office31 dataset is a mainstream image benchmark for transfer learning, which contains 31 categories with a total of 4,652 images. There are three different domains (Amazon, Webcam, and DSLR) in this dataset. We select Webcam and DSLR, which are data-hungry, as the target domain, and use Amazon, which is rich in data, as the source domain. Then we construct two datasets: Office31 (A$\rightarrow$W) and Office31 (A$\rightarrow$D). For the two datasets, all source domain samples are used for training, and each category in the target domain randomly selects 3 samples for training, and the remaining samples are further divided into the validation set and test set in equal proportions.

• FB: FB (Facebook100) datasets are social networks of 100 different universities in the United States. Each sample (students or faculties) have six attributes (e.g., gender, major, and dorm), and the goal is to predict Node identity flags. We view the social network of each university as a different domain, and select two of the university’s social networks and put them together to form a cross-network dataset in $RD_{intra}$ scenarios. In this paper, we select four universities, and construct two cross-network datasets: FB (Hamilton$\rightarrow$Caltech) and FB (Howard$\rightarrow$Simmons).

• Company: Company dataset is a company investment network with 10641 real-world companies in China. The node attributes are the company's financial attributes. We regard listed companies as source domain and unlisted companies as target domain and the goal is to predict the risk status of non-listed companies. We complement missing features of non-listed companies by mean value of neighbors (we do not consider the feature missing problems in the original paper ). And we use this dataset in the $RD_{intra-and-inter}$ scenario.

• Sync-$UD$/Sync-$RD_{intra}$/Sync-$RD_{intra-and-inter}$: We also construct a synthetic dataset for three scenarios of GKT by randomly sampling points of source and target domains from two distinct Multivariate Gaussian distributions. As shown in the Fig. \ref{fig:domain_shift_visual} (e), samples of source and target domains in this dataset are designed to have distinct conditional distribution and marginal distribution to validate the motivations described in Sec. \ref{sec:motivation}. We use this dataset in all three scenarios, and we randomly add edges with fixed homophilous ratio of 70% for Sync-$RD_{intra}$ and Sync-$RD_{intra-and-inter}$.

Training Bridged-GNN

First, unzip the datasets.zip file into the datasets folder, and the code can be found at the ./Bridged-GNN folder.

Then, we provide the demo file (run.sh) to train a Bridged-GNN model:

sh run.sh

Specifically, the training process are divided into two steps:

• Step-1: to learn the bridged-graph first (office_amazon2dslr datasets for example)

  Run

  python main_bridged_graph.py --hidden_dim 128 --num_epoch 400 --start_eval_epoch 300 --epsilon 0.5 --k_within 3 --k_cross 20  --seed 0  --save --dataset_name office_amazon2dslr --version v2  --check_within --check_cross 

Then the learned bridged-graph will be saved into the ./data_bridged_graph folder, and the model parameters will be saved into the ./ckpt folder

• Step-2: to conduct knowledge transfer on the learned bridged-graph with GNN (here we use KT-GNN):

  Run

  python main_graph_knowledge_transfer.py --num_layer 2 --hidden_dim 64 --path_data ../data_bridged_graph/office_amazon2dslr_bridged_graph.dat --to_undirected