Skip to content

readme.md

Latest commit

 

History

History
111 lines (77 loc) · 3.61 KB

readme.md

File metadata and controls

111 lines (77 loc) · 3.61 KB

A guidance for script

Here is the base script containing three terms.

bash script/run.sh DATASET BACKBONE SAMPLING

The first is dataset, which indicates the source and target of the transfer learning. We provide a specific guide on how to set the dataset in the following.

The second is backbone, which can be chosen from gcn, gat, sage, sgc. More architectures can be implemented under model.py.

The last term is SAMPLING involving the subgraph sampling method in our proposed SP. we provide two sampling functions including k_hop (k-hop sampler) and rw (random walk sampler), corresponding to SP and SP++, respectively.

Here are some examples.

bash script/run.sh acm_dblp gcn k_hop
bash script/run.sh dblp_acm gcn k_hop
bash script/run.sh arxiv_1_arxiv_5 gcn rw
bash script/run.sh arxiv_3_arxiv_5 gcn rw

How to set Dataset in script?

Citation Network

There are two datasets: ACM and DBLP. We test the one-to-one transfer setting.

  • Transfer From ACM to DBLP: acm_dblp
  • Transfer From DBLP to ACM: dblp_acm

Airport Network

There are three datasets: USA, Europe, Brazil. We test the one-to-one transfer setting.

  • Transfer From USA to Europe: usa_europe
  • Transfer From USA to Brazil: usa_brazil
  • Transfer From Europe to USA: europe_usa
  • Transfer From Europe to Brazil: europe_brazil
  • Transfer From Brazil to USA: brazil_usa
  • Transfer From Brazil to Europe: brazil_europe

Twitch Network

There are six networks collected from different countries. We test the one-to-multi transfer learning performance. Particularly, the knowledge is transferred from DE to the remaining datasets, i.e., EN, ES, FR, PT, RU.

  • Transfer From DE to EN: de_en
  • Transfer From DE to ES: de_es
  • Transfer From DE to FR: de_fr
  • Transfer From DE to PT: de_pt
  • Transfer From DE to RU: de_ru

Arxiv Network

We test the temporal dynamic distribution shift on the dataset. Specifically, this is a citation network where papers are published from 2005 to 2020. We consider five splits: Time 1 (2005 - 2007), Time 2 (2008 - 2010), Time 3 (2011 - 2014), Time 4 (2015 - 2017), Time 5 (2018 - 2020). We transfer the knowledge from the previous four datasets to the last one.

  • Transfer From T1 to T5: arxiv_1_arxiv_5
  • Transfer From T2 to T5: arxiv_2_arxiv_5
  • Transfer From T3 to T5: arxiv_3_arxiv_5
  • Transfer From T4 to T5: arxiv_4_arxiv_5

Additionally, we also consider one domain adaptation setting (Degree).

  • arxiv_arxiv_0

Elliptic Network

This is another temporal dynamic graph. Please directly run the specific script.

bash script/run_elliptic.sh elliptic BACKBONE SAMPLING

The example is

bash script/run_elliptic.sh elliptic gat rw

Facebook Network

This is a social dataset collected on Facebook, which consists of 14 networks. Please run the specific script.

bash script/run_fb.sh DATASET BACKBONE SAMPLING

The model is trained on 3 graphs, validated on 2 graphs, and test on 1 graph. Here is an example

bash script/run_fb.sh facebook_1_2_3_10 gcn k_hop

The dataset looks like facebook_1_2_3_10 where different numbers indicate different graphs. We use the graphs with first three indices as training graphs and the last one as testing. In this example, 1, 2, 3 are training graphs and 10 is testing graph. Note that graphs with indices 13 and 14 are used as validations for all of settings.

Here is the specific mapping:

  1. Johns Hopkins55
  2. Caltech36
  3. Amherst41
  4. Bingham82
  5. Duke14
  6. Princeton12
  7. WashU32
  8. Brandeis99
  9. Carnegie49
  10. Penn94
  11. Brown11
  12. Texas80
  13. Cornell5
  14. Yale4