Comparing three robust procedures for CANDECOMP/PARAFAC estimation

CANDECOMP/PARAFAC aims to identify the true components underlying data with a trilinear configuration. The search for a unique solution is not always an easy task, as degeneracies may occur. The presence of outlier contamination further complicates the matter by requiring the implementation of robust procedures. The most used robust approach RALS is based on the iterative repetition of the standard alternating least squares algorithm, which is known to be slow and vulnerable to over-factoring, collinearity, and bad initial values. Here the faster and stable robust alternative R-INT1, based on the SWATLD-ALS integrated scheme INT-1, is implemented. Its performance is tested against ALS, R-ALS, and R-INT2 (built on INT-2, an ATLD-ALS procedure already proposed in the literature). Performance is assessed in a simulation study with varied levels of outlier contamination.