Data Synthesis for Testing Black-Box Machine Learning Models

Saha, Diptikalyan; Aggarwal, Aniya; Hans, Sandeep

doi:10.1145/3493700.3493704

“…Variational auto-encoders and generative adversarial networks are examples of neural methods [Fa20]. There are various tools available that implement statistical and/or neural data synthesis methods for realistic tabular data, "AIT-EST" [SAH22] or the "Synthetic Data Vault" for instance [Pa16].…”

Section: Related Workmentioning

confidence: 99%

“…To synthesize learning data specific to the Moodle Learning Management System, we follow the general approach for data synthesis suggested in [EMH20], combined with suggestions from [Fa20] and [SAH22]. [EMH20] break down a general data synthesis process into three basic steps.…”

Section: Approachmentioning

confidence: 99%

Data Synthesis for Fairness Audits of Learning Analytics Algorithms

Fernsel¹,

Simbeck²

2022

Angewandte Forschung in Der Wirtschaftsinformatik 2022

1

0

View full text Add to dashboard Cite

The purpose of methods of fairness auditing is to uncover to what extent Learning Analytics algorithms are fair. Fairness auditing methods often rely on pre-existing test data. In the context of Learning Analytics auditing, learning data is needed for testing. However, learning data might not be available (in large quantities) due to privacy concerns. Our poster shares our findings on how relational data for fairness audits of Learning Analytics systems can be synthesized from little pre-existing data, using the most promising available data synthesizers.

show abstract

“…Variational auto-encoders and generative adversarial networks are examples of neural methods [Fa20]. There are various tools available that implement statistical and/or neural data synthesis methods for realistic tabular data, "AIT-EST" [SAH22] or the "Synthetic Data Vault" for instance [Pa16].…”

Section: Related Workmentioning

confidence: 99%

“…We will evaluate whether different learning styles are represented in the synthetic data and whether the synthetic data can be used for auditing the dropout prediction model. To synthesize learning data specific to the Moodle Learning Management System, we follow the general approach for data synthesis suggested in [EMH20], combined with suggestions from [Fa20] and [SAH22]. [EMH20] break down a general data synthesis process into three basic steps.…”

Section: Our Contributionmentioning

confidence: 99%