Adaptive Divergence-Based Non-Negative Latent Factor Analysis of High-Dimensional and Incomplete Matrices From Industrial Applications
Ye Yuan, Xin Luo, MengChu Zhou
Abstract
<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">H</u> igh- <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</u> imensional and <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">I</u> ncomplete (HDI) data are commonly seen in various big-data-related applications concerning the inherent non-negativity interactions among numerous nodes. A <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N</u> on-negative <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">L</u> atent <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</u> actor <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</u> nalysis (NLFA) model performs efficient representation learning to such HDI data. However, existing NLFA models all adopt a static divergence metric like Euclidean distance or <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">α</i> - <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">β</i> divergence to build its learning objective, which evidently restricts its scalability in representing HDI data from different domains. Aiming at addressing this critical issue, this study proposes an <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</u> daptive <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</u> ivergence-based <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N</u> on-negative <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">L</u> atent-factor-analysis (ADNL) model with three-fold ideas: a) generalizing the objective function with the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">α</i> - <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">β</i> -divergence to expand its potential of representing various HDI data; b) facilitating a smooth non-negative bridging function to connect the optimization variables with output latent factors for keeping non-negativity; and c) making the divergence parameters adaptive through position-transitional particle swarm optimization, thereby facilitating adaptive divergence in the learning objective to achieve high scalability. Empirical studies on six HDI datasets from real applications demonstrate that an ADNL model outperforms the state-of-the-art models in both estimation accuracy and computational efficiency for missing data of an HDI matrix.