Novel Algorithm for Incremental L1-Norm Principal-Component Analysis

Abstract

L1-norm Principal-Component Analysis (L1-PCA) has been shown to exhibit sturdy resistance against outliers among the processed data. In this work, we propose L1-IPCA: an algorithm for incremental L1-PCA, appropriate for big-data and streaming-data applications. The proposed algorithm updates the calculated L1-norm principal components as new data points arrive, conducting a sequence of computationally efficient bit-flipping iterations. Our experimental studies on subspace estimation, image conditioning, and video foreground extraction illustrate that the proposed algorithm attains remarkable outlier resistance at low computational cost.