<p>Graphs serve as a vital data representation model across various real-world applications including biological research and social network analysis. Many machine learning approaches need a high ability to effectively learn and extract insights from graph structures. Graph embedding plays a crucial role in transforming non-Euclidean feature spaces into low-dimensional structured representations, so it can be easily utilized by machine learning methods. However, the complexity of graph representation learning, especially in node classification tasks arises from the intricate interactions among the labelled information, node attributes, significant features, topological structures, and node types. Current graph convolutional networks (GCNs) often struggle with missing node attribute issues and exhibit inefficiencies during the propagation of information. Hence, an efficient node embedding technique is introduced in this research work to overcome the limitations of the classical techniques. Initially, the graph data needed for performing the node embedding process is obtained from different benchmark sources. Next, the random walk regularization-based node representation learning is introduced to learn the node attributes as it helps to analyse the potential node representation. The random walk model is used in this work, as it has the efficiency to collect and convert the geometric structure into a structured sequence. Then, an efficient network named Adaptive Graph Recurrent Autoencoder with Attention Network (Ada-GRAE-AN) is implemented to execute the node embedding procedures. The developed framework utilizes the attention mechanism to distinguish the importance of neighbouring nodes. Moreover, the parameters in Ada-GRAE-AN are optimized using an Improved Random Parameter-based Piranha Foraging Optimization Algorithm (IRP-PFOA). The major objective of the developed framework is to improve the node embedding with graph labelling information, including the structural information in the node embedding process and enhance the downstream task performance. The developed framework enhances the task performance by classifying the nodes and also predicting the links. Further, various experimental validations are performed in the developed framework over the classical techniques to ensure its effectiveness.</p>