A0844
Title: Optimal semi-supervised subsampling for softmax regression
Authors: Ai-Lin Wang - National Taiwan University and Academia Sinica (Taiwan) [presenting]
Frederick Kin Hing Phoa - Academia Sinica (Taiwan)
HaiYing Wang - University of Connecticut (United States)
Abstract: Missing label information presents a major challenge for optimal subsampling methods, as most require access to the full set of response labels to construct sampling probabilities. A- and L-optimal semi-supervised subsampling strategies are investigated for softmax regression, focusing on deriving optimal subsampling distributions under the baseline constraint. Furthermore, the semi-supervised constraint-invariant approach is explored by minimizing the asymptotic mean squared prediction error, enabling the construction of subsampling probabilities that are robust to model constraint choices. Results demonstrate that effective semi-supervised subsampling is possible even with partial label information.
