Cloud computing has emerged as the feasible paradigm to satisfy the computing requirements of high-performance applications by an ideal distribution of tasks to resources. But, it is problematic when attaining multiple scheduling objectives such as throughput, makespan, and resource use. To resolve this problem, many Task Scheduling Algorithms (TSAs) are recently developed using single or multi-objective metaheuristic strategies. Amongst, the TS based on a Multi-objective Grey Wolf Optimizer (TSMGWO) handles multiple objectives to discover ideal tasks and assign resources to the tasks. However, it only maximizes the resource use and throughput when reducing the makespan, whereas it is also crucial to optimize other parameters like the utilization of the memory, and bandwidth. Hence, this article proposes a hybrid TSA depending on the linear matching method and backfilling, which uses the memory and bandwidth requirements for effective TS. Initially, a Long Short-Term Memory (LSTM) network is adopted as a meta-learner to predict the task runtime reliability. Then, the tasks are divided into predictable and unpredictable queues. The tasks with higher expected runtime are scheduled by a plan-based scheduling approach based on the Tuna Swarm Optimization (TSO). The remaining tasks are backfilled by the VIKOR technique. To reduce resource use, a particular fraction of CPU cores is kept for backfilling, which is modified dynamically depending on the Resource Use Ratio (RUR) of predictable tasks among freshly submitted tasks. Finally, a general simulation reveals that the proposed algorithm outperforms the earlier metaheuristic, plan-based, and backfilling TSAs.