Resource allocation in Kubernetes involves efficiently distributing Central Processing Unit (CPU), memory, and other resources among containers to ensure optimal performance and scalability. Resource allocation in Kubernetes is that inaccurate requests and limits can lead to resource wastage or application performance degradation. In this manuscript, optimization of resource allocation in Kubernetes-based Machine Learning (ML) systems using predictive scaling (ORAK-MLS-PS) is proposed. Firstly, input data is gathered from Kubernetes Resource and Performance Metrics Allocation dataset. Then, data is pre-processed utilizing Continuous-Discrete Derivative-Free Extended Kalman Filter (CD-DFEKF), which is used for normalization. Then the pre-processed data are given to Progressive Graph Convolutional Networks (PGCN) to predict resource allocation in Kubernetes by forecasting workload demands to enhance performance and efficiency. The Leaf in Wind Optimization Algorithm (LWO) is used to optimize the weight factors of PGCN. The proposed ORAK-MLS-PS approach is implemented in Python. The proposed ORAK-MLS-PS achieves an accuracy of 98.9%, outperforming the other existing methods HERC-KML-EC 75.4%, KBS-ERA-CW 78.7%, and AHS-KC-ANN 85.2% with existing methods, like HERCULE: high-efficiency resource coordination using Kubernetes and machine learning in edge computing for improved QoS and QoE (HERC-KML-EC), a Kubernetes-driven system for effectual resource allocation in containerized workflow (KBS-ERA-CW) and adaptive horizontal scaling in Kubernetes clusters with ANN-driven load forecasting (AHS-KC-ANN).