Datasets
Standard Dataset
Fair-B-PG
- Citation Author(s):
- Submitted by:
- Manjish Pal
- Last updated:
- Sun, 03/10/2024 - 10:06
- DOI:
- 10.21227/dh84-y945
- License:
21 ViewsAbstract
Weconsiderfivebenchmarkdatasets-Pokec-z,NBA,
Political-Blogs, Cora and Twitter which are predominantlyWeconsiderfivebenchmarkdatasets-Pokec-z,NBA,We consider five benchmark datasets - Pokec-z, NBA, Political-Blogs, Cora and Twitter which are predominantly used in the fair link prediction literature \cite{dong2022fairness}. Since Political-Blogs, Cora and Twitter datasets have a single sensitive attribute we compare the fairness vs. link prediction performance for that case whereas for Pokec-z and NBA datasets we also demonstrate comparison for the case of multiple sensitive attributes. Political-Blogs, Cora and Twitter datasets consist of a single sensitive attribute and hence utility and fairness results could not be obtained for intersectional groups. However, for Pokec and NBA which consists of multiple sensitive attributes, both intersectional and overlapping group fairness results along with utility are reported.
Political-Blogs, Cora and Twitter which are predominantly
used in the fair link prediction literature [10]. Political-Blogs,
Cora and Twitter datasets consist of a single sensitive attribute
and hence utility and fairness results could not be obtained for
intersectional groups. However, for Pokec and NBA which
consists of multiple sensitive attributes, both intersectional
and overlapping group fairness results along with utility are
reported. We assume each node to be associated with a set
of features and an (undirected) edge represents a relationship
among a pair of nodes.used in the fair link prediction literature [10]. Political-Blogs,
Cora and Twitter datasets consist of a single sensitive attribute
and hence utility and fairness results could not be obtained for
intersectional groups. However, for Pokec and NBA which
consists of multiple sensitive attributes, both intersectional
and overlapping group fairness results along with utility are
reported. We assume each node to be associated with a set
of features and an (undirected) edge represents a relationship
among a pair of nodes.
We consider five benchmark datasets - Pokec-z, NBA, Political-Blogs, Cora and Twitter which are predominantly used in the fair link prediction literature Political-Blogs, Cora and Twitter datasets consist of a single sensitive attribute and hence utility and fairness results could not be obtained for intersectional groups. However, for Pokec and NBA which consists of multiple sensitive attributes, both intersectional and overlapping group fairness results along with utility are reported. We assume each node to be associated with a set of features and an (undirected) edge represents a relationship among a pair of nodes. Further details related to the datasets are provided in Table attached.
Documentation
| Attachment | Size |
|---|---|
 table.pdf | 103.39 KB |
