Self-Evolving Autonomous Software Architectures Using Large-Scale Graph Neural Networks and Real-Time Big Data Feedback Loops for Economic Optimization and Cost-Efficient Resource Allocation

The rapid growth of cloud-native, microservice-based, and distributed computing environments has exposed the limits of conventional software architectures that rely on static rules, manually configured resource policies, and human-driven adaptation. These limitations often produce resource overprovisioning, higher operational costs, delayed failure recovery, and inefficient use of computational capacity. This study proposes a Self-Evolving Autonomous Software Architecture (SEASA) that combines large-scale Graph Neural Networks (GNNs) with real-time big-data feedback loops to enable continuous architectural learning, economic optimization, and cost-efficient resource allocation. The framework represents software ecosystems as dynamic heterogeneous graphs, where services, containers, databases, nodes, and infrastructure resources are modelled as interconnected entities, while dependencies, communication patterns, costs, and data flows are modelled as evolving edges. Runtime data were collected from Microsoft Azure Cloud workload traces, Google Cluster Workload Traces, and a synthetic microservice benchmark containing 500 services and 2.8 million interactions. Graph Convolutional Network, GraphSAGE, Graph Attention Network, and Temporal Graph Neural Network models were evaluated for anomaly detection, workload forecasting, architectural-state prediction, and autonomous adaptation. The TGNN-based framework achieved 97.4% accuracy, 96.8% precision, 96.1% recall, and a 96.4% $F_1$-score, outperforming baseline models. It reduced adaptation latency by 38.7%, resource consumption by 24.5%, cloud operating costs by 21.8%, and overprovisioning by 27.3%. It also improved failure recovery, allocation efficiency, and service-level-objective compliance. These findings show that temporal graph intelligence and real-time economic feedback can transform software architectures into predictive, self-directed, and cost-aware computing ecosystems.

The SEASA framework models the software ecosystem as a dynamic heterogeneous graph $G_t=(V_t,E_t)$ at time $t$, where vertices $V_t$ represent services, containers, databases, nodes, and infrastructure resources, and edges $E_t$ encode dependencies, communication patterns, costs, and data flows. The graph evolves over time, capturing the dynamic nature of cloud environments. To process this graph, the framework employs Graph Neural Networks (GNNs) that learn representations of nodes and edges. Specifically, Graph Convolutional Networks (GCNs), GraphSAGE, Graph Attention Networks (GATs), and Temporal Graph Neural Networks (TGNNs) are evaluated. The TGNN is particularly suited for capturing temporal dependencies, using mechanisms such as recurrent units or temporal convolutions to update node embeddings over time. The GNN outputs are used for multiple tasks: anomaly detection, workload forecasting, architectural-state prediction, and autonomous adaptation. A real-time big data feedback loop continuously ingests runtime data from cloud workload traces (Microsoft Azure, Google Cluster) and a synthetic microservice benchmark (500 services, 2.8 million interactions). This data is used to update the graph and retrain the models, enabling self-evolution. The economic optimization component formulates cost-efficient resource allocation as an optimization problem: minimize operational cost subject to service-level objectives (SLOs). The GNN predictions inform a decision-making module that adjusts resource allocations, scaling policies, and failure recovery actions. The framework is evaluated against baseline models (static rules, manual policies) using metrics such as accuracy, precision, recall, $F_1$-score, adaptation latency, resource consumption, cloud operating costs, overprovisioning, failure recovery time, allocation efficiency, and SLO compliance. The architecture is depicted in the Mermaid diagram below: