Effective Network Partial Recovery System Using the Memorized Random Rescue Optimized Deep Convolutional Neural Network for Recommendation Model Training

Abstract

Fault tolerance is a crucial part of a stream processing system that ensures consistent data analysis even after failures or breakdowns in the system utilizing the rollback and checkpointing mechanism. However, due to the increasing frequency of failures, delay, the variability of fault checkout overhead, and checkout time consumption, the systems will not meet the Quality of Service constraints introducing issues in the system's performance. Additionally, the performance of the systems seems to be more difficult due to the high-performance computing platform's expanding physical size. To overcome these limitations, a Memorized Random Rescue Optimization-based deep Convolutional Neural Network (MONN) partial recovery model for the training recommendation system is developed and optimized. The deep CNN parameters are tuned utilizing the Memorized Random Rescue Optimization (MRRO) algorithm to optimally select the checkpoint interval which is loaded from the broken point and returned to the classifier to enhance the partial recovery. By allowing non-failed systems to continue before loading checkpoints whenever a system breaks down throughout training, lowering failure-related overheads, and reducing the trained model's recovery period, the MONN model reduces the reliability conditions. The MONN model, which uses a partial recovery training approach to shorten the training time while maintaining the target level of model accuracy, is proposed as a result of the analysis. The efficiency of the developed MONN model is evaluated in terms of metrics which are reported as 94.393% accuracy,72.956% precision,76.817% recall, and 7.424 s recovery time utilizing the database 1. The developed model attained 95.778% accuracy, 6.585 s recovery time, 95.069% precision, and 78.028% recall utilizing database 2, concerning the 90% of training. Furthermore, the MONN model attained an accuracy of 93.960%, a recovery time of 3.628 s, a precision of 94.678%, and, a recall of 77.444% utilizing dataset 3 for the 90% of training reveals the superiority of the developed model compared with other conventional techniques.