In the context of edge computing and the Internet of Things, where computational resources are deployed closer to end users to reduce latency and bandwidth usage, dynamic resource management becomes critical. The Horizontal Pod Autoscaler(HPA) in Kubernetes, while effective in cloud environments, lacks the intelligence to account for edge-specific constraints such as request redirection, dynamic network latency, and fine-grained geospatial demand. This leads to suboptimal autoscaling and inefficient resource usage in edge clusters. Other existing autoscaling approaches primarily rely on static thresholds or simplistic CPU/memory metrics, which are inadequate for latency-sensitive edge services. This study introduces MATRIX, a machine learning-driven extension to Kubernetes HPA, tailored for edge environments. By incorporating features such as request history, user geolocation, server-to-server communication delay, pod-specific redirection ratios, and real-time edge node load metrics, MATRIX predicts which edge nodes should replicate a given pod to ensure service continuity and efficiency. Unlike existing threshold-based approaches, our model learns optimal scaling strategies through supervised learning and provides intelligent decisions that adapt to real-time conditions. Experimental evaluations on a simulated Kubernetes-based edge infrastructure demonstrate significant improvements in latency reduction, bandwidth optimization, and workload distribution compared to the standard HPA. MATRIX is especially suitable for latency-sensitive and highly distributed applications such as smart healthcare, autonomous vehicles, and immersive virtual environments.