High-Efficiency System Control for Latent Factor Analysis via PID-incorporated SGD Algorithm

High-dimensional and incomplete (HDI) matrix is widely applied in representing complicated interactions among diverse nodes in large-scale industrial applications. The latent factor analysis (LFA) model is accomplished in digging out serviceable latent characteristics of nodes from HDI matrix based on the stochastic gradient descent (SGD) algorithm. However, the SGD algorithm merely depends on the current information of a single instance when updating the LF matrices, while ignoring the historical information and underlying connections between instances, making the resulting model converge tardily and lack systematicness. To tackle this crucial issue, this paper proposes an PSLF model according to constructs a PID-accelerate SGD (PAS) algorithm by combining correlated instances and refining learning error with current, historical, and future trend information by proportional-integral-derivative (PID) mechanism. Furthermore, the comparative experiments on two HDI matrices demonstrate the superiority of the PSLF model in representation capacity on the HDI matrix.