Event Title
Deep Metric Learning to Evaluate Student Performance on Standardized Tests
Date of Publication
2-3-2021
Document Type
Presentation
Abstract
We propose a triplet network design which generates vectors representing student understanding. The model was trained on a sample of students (n = 393,609) and questions (n = 54) from the 2017 seventh grade math STAAR test. The model predicts student results more accurately than a randomly initialized model, by a factor of 1.96. Triplet learning networks are known to provide vectors with a useful distance metric, even when given discrete data (Hoffer & Ailon, 2014), which offers opportunities for novel analysis methods. For example, by using a clustering algorithm, educators could more precisely target instruction to students.
Keywords
Standarized Tests, Performance
Persistent Identifier
http://hdl.handle.net/10950/2950
Deep Metric Learning to Evaluate Student Performance on Standardized Tests
We propose a triplet network design which generates vectors representing student understanding. The model was trained on a sample of students (n = 393,609) and questions (n = 54) from the 2017 seventh grade math STAAR test. The model predicts student results more accurately than a randomly initialized model, by a factor of 1.96. Triplet learning networks are known to provide vectors with a useful distance metric, even when given discrete data (Hoffer & Ailon, 2014), which offers opportunities for novel analysis methods. For example, by using a clustering algorithm, educators could more precisely target instruction to students.