.png){kind=icon}
.png){kind=icon}
Return
Lessons Learned: Can Machine Learning Model Expose Dataset Contamination?
.png){kind=icon}
PRE
AI
DOI：10.1175/AIES-D-25-0030.1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Dataset contamination is a typical challenge when dealing with large quantities of multidimensional data and is expected to impact any data-specific applications. While several error diagnostic methods exist to assess the datasets for biases, some errors might go unnoticed by developers and users. In this paper, we document our experience detecting errors in the dataset that were missed by other validation methods but identified by analyzing machine learning model predictions trained on contaminated datasets. Specifically, we used simulated observations of a future space-borne passive microwave sensor, the Second-Generation EUMETSAT Polar System Microwave Imager, to develop a convolutional neural network-based model to predict total column water vapor. An oversight in time matching during the data creation led to contaminating a small portion of the dataset. Although the simulated data did not show any signs of the time mismatch and the overall statistical scores of predictions were good, artifacts identified in the model output led to an in-depth analysis of individual input orbits revealing high root-mean-square error values for the contaminated samples, highlighting the erroneous points in the simulated brightness temperature data. This work also evaluates the impact of data contamination on data-driven frameworks across different proportions of erroneous samples while highlighting the importance of data curation.
Keywords:
Atmosphere
Error analysis
Neural networks
Machine learning
Deep learning
Journal
A
IF:
0
Papers:
63
Citations:
0