Multi-perspective cross-indicator correlation modeling for health status prediction of VR software

In response to the inherent long-term and short-term degradation characteristics exhibited by the time-series data of health status indicators for VR software, a novel VR software health status prediction method (named MPCI) based on Multi-Perspective Cross-Indicator association modeling is proposed. Firstly, this method introduces a hybrid algorithm combining category balancing and synthetic sampling, which integrates undersampling and oversampling techniques to address the issue of data category imbalance. Secondly, a modeling approach with a dual-branch interactive architecture is proposed to comprehensively capture the temporal correlations of both local and global features, thereby fully extracting the temporal information of the indicators. Finally, by incorporating a Long Short-Term Memory (LSTM) component combined with Self-Attention, the method learns the non-linear mapping relationships between each indicator. Additionally, through an improved Stacked Denoising AutoEncoder (SDAE) component, it extracts the spatial information of pairwise interaction relationships among the indicators. Using Run-to-Failure data collected from Android and OpenStack systems, the proposed method is compared with four baseline models, demonstrating its superiority in prediction accuracy. For the response time and idle memory test sets, across multi-step prediction horizons ranging from 1 to 9, the proposed method yields mean absolute percentage error values within the ranges of [0.0611–0.0979] and [0.0073–0.0207], respectively. These results indicate that this study provides more reliable decision support for predictive maintenance of VR software health status and can effectively optimize proactive recovery scheduling strategies. © 2026 Published by Elsevier Inc.