| Abstract |
Deconvolution is the non-parametric estimation of the distribution of a target variable. If the variable cannot be measured directly, it has to be reconstructed from contaminated observations. The Dortmund Spectrum Estimation Algorithm (DSEA) formulates this problem as a multinomial classification task. It may be more appropriate to use regression methods if the target variable of interest is continuous. We present CSEA, an extension of DSEA for continuous deconvolution. CSEA uses a combination of regression analysis and kernel density estimation. Existing methods revolve around binning the data and fitting functions to the probabilities assigned to the bins. The consideration of all predictions leads to a more granular solution and avoids discretization errors. Our development of a robust parameterization allows CSEA to outperform its discrete counterpart in the experiments conducted.
|
