We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Bias Reduction via End-to-End Shift Learning: Application to Citizen Science

Abstract: Citizen science projects are successful at gathering rich datasets for various applications. However, the data collected by citizen scientists are often biased --- in particular, aligned more with the citizens' preferences than with scientific objectives. We propose the Shift Compensation Network (SCN), an end-to-end learning scheme which learns the shift from the scientific objectives to the biased data while compensating for the shift by re-weighting the training data. Applied to bird observational data from the citizen science project eBird, we demonstrate how SCN quantifies the data distribution shift and outperforms supervised learning models that do not address the data bias. Compared with competing models in the context of covariate shift, we further demonstrate the advantage of SCN in both its effectiveness and its capability of handling massive high-dimensional data.
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Cite as: arXiv:1811.00458 [cs.LG]
  (or arXiv:1811.00458v4 [cs.LG] for this version)

Submission history

From: Di Chen [view email]
[v1] Thu, 1 Nov 2018 15:54:30 GMT (6592kb,D)
[v2] Fri, 2 Nov 2018 21:32:48 GMT (6592kb,D)
[v3] Tue, 13 Nov 2018 17:36:08 GMT (6425kb,D)
[v4] Wed, 14 Nov 2018 05:35:37 GMT (6425kb,D)

Link back to: arXiv, form interface, contact.