FEDERATED L1-AND-L2-NORM-ORIENTED LATENT FACTOR MODEL for ANALYZING HIGH-DIMENSIONAL and INCOM-PLETE DATA

High-dimensional and incomplete (HDI) data are extremely prevalent in numerous big-data applications. Latent Feature Analysis (LFA), as a classic representation learning method, with the help of low-rank embedding technology, can unearth the latent and valuable knowledge hidden in HDI data. However, most existing LFA-based models use fixed and exclusive L₂-norm-oriented loss along with centralized strategies, risking privacy leakage. Recognizing this, this paper innovatively introduces a Federated L₁-and-L₂-norm-oriented Latent Feature Analysis (FedLF) model, whose main ideas are twofold: 1) employing L₁-and-L₂-norm-oriented loss to construct two variant Autoencoders, each of which adopts federated learning to protect privacy, and 2) organically aggregate all variants by using a skillfully tailored adaptive weighting strategy. In this way, our FedLF can ensure an all-inclusive and impartial representation of HDI data while ensuring data privacy. Comprehensive experiments on two real-world HDI matrices clearly confirm that FedLF surpasses the performance of leading contemporary models.