International Journal of Scientific Research and Engineering Development

We propose the Dynamic Hierarchical Attention Graph (DHAG), a novel neural architecture that dynamically constructs and updates hierarchical graph representations for real-time feature importance analysis in AI-computer science fusion applications. Traditional approaches often rely on static feature selection or fixed windowing techniques, which lack adaptability to evolving data streams; the proposed method addresses this limitation by integrating dynamic graph attention networks with temporal graph embedding techniques. The architecture operates through three core modules: an adaptive attention mechanism that computes feature importance scores, a dynamic hierarchical graph construction process that clusters nodes based on these scores, and a memory-augmented temporal graph neural network that captures long-term dependencies. The hierarchy is pruned using expander graph properties to maintain sparsity, while rotary positional embeddings ensure temporal alignment. Moreover, the system replaces conventional static feature selection with a data-driven hierarchy, enabling context-aware embeddings that adapt to real-time input patterns. The implementation employs multi-head attention with residual connections and a graph transformer architecture, guaranteeing both scalability and efficiency. Experiments demonstrate that DHAG significantly reduces inference latency by pruning irrelevant subgraphs while maintaining high accuracy. This work bridges the gap between dynamic graph representation learning and real-time feature importance analysis, offering a scalable solution for emerging applications in AI-driven systems. The results highlight its potential to outperform static baselines in scenarios requiring adaptive, context-sensitive decisionmaking.