The popular Alternating Least Squares (ALS) algorithm for tensor decomposition is extremely efficient, but often converges to poor local optima, particularly when the weights of the factors are non-uniform. We propose a modification of the ALS approach that is as efficient as standard