Datasets:

Modalities:
Tabular
Formats:
csv
ArXiv:
Libraries:
Datasets
pandas
License:
mvladimirova commited on
Commit
941b091
1 Parent(s): ae1fe47

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +5 -4
README.md CHANGED
@@ -73,10 +73,11 @@ Readers should not interpret summary statistics of this dataset as ground truth
73
 
74
  ## Challenges
75
 
76
- The first challenge comes from handling the different types of data that are common in tables, the mixed-type columns: there are both numerical and categorical features that have to be embedded [Gorishniy et al., 2021, 2022, Grinsztajn et al., 2022, Shwartz-Ziv and Armon, 2022, Matteucci et al., 2023].
77
- In addition, some of the features have long-tail phenomenon and products have popularity bias. Our datasets contains more than 1,000,000 lines, while current high-performing models are under-explored in scale, e.g. the largest datasets in Grinsztajn et al. [2022] are only 50,000 lines, while in Gorishniy et al. [2021, 2022] only one dataset surpasses 1,000,000 lines.
78
- Additional challenge comes from strongly imbalanced data: the positive class proportion in our data is less than 0.007 that leads to challenges in training robust and fair machine learning models [Jesus et al., 2022, Yang et al., 2024]. In our dataset there is no significant imbalances in demographic groups users regarding the protected attribute (both genders are sub-sampled with 0.5 proportion, female profile users were shown less job ad with 0.4 proportion and slightly less senior position jobs with 0.48 proportion), however, there could be a hidden effect of a selection bias.
79
- This poses a problem in accurately assessing model performance [van Breugel et al., 2024].
 
80
  More detailed statistics and exploratory analysis are referred to the supplemental material of the associated paper linked below.
81
 
82
  ## Metrics
 
73
 
74
  ## Challenges
75
 
76
+ The first challenge comes from handling the different types of data that are common in tables, the mixed-type columns: there are both numerical and categorical features that have to be embedded.
77
+ In addition, some of the features have long-tail phenomenon and products have popularity bias. Our datasets contains more than 1,000,000 lines, while current high-performing models are under-explored in scale.
78
+ Additional challenge comes from strongly imbalanced data: the positive class proportion in our data is less than 0.007 that leads to challenges in training robust and fair machine learning models.
79
+ In our dataset there is no significant imbalances in demographic groups users regarding the protected attribute (both genders are sub-sampled with 0.5 proportion, female profile users were shown less job ad with 0.4 proportion and slightly less senior position jobs with 0.48 proportion), however, there could be a hidden effect of a selection bias.
80
+ This poses a problem in accurately assessing model performance.
81
  More detailed statistics and exploratory analysis are referred to the supplemental material of the associated paper linked below.
82
 
83
  ## Metrics