|
Abstract:
|
Matrix completion, a process of using a low rank approximation to impute missing values in a data matrix, offers a general and powerful framework for imputation. It has received significant attention over the last decade and a plethora of new methods for faster, more accurate convergence have been developed. Two main limitations remain restricting its implementation in statistical practice - choosing an appropriate rank and scoring new observations without needing to recalculate the entire model (i.e. performing streaming matrix completion). These limitations result in matrix completion remaining a relatively esoteric imputation method, when often the generality of a low rank approximation fits a variety of data well. We suggest a way to address both problems, which allows for a fully automated streaming matrix completion algorithm. While fine tuning options are available, this algorithm requires at a minimum only a single click from the user to fit a streaming imputation model with error estimates generated by cross-validation
|
