Methods for assessing embedding quality quantify the similarity between the neighbourhood structure in the original higher dimension and the embedded dimension of individual data points. These also quantify the similarity of the data’s global structure in higher and embedded dimensions. These methods are important when evaluating the efficacy of a dimensionality reduction (DR) method because DR methods transform data from a higher dimensional to a lower dimensional space. The most efficient EQA method, manifold embedding quality assessment (MEQA), uses transformation and statistical analysis to measure local and global structure similarity. The transformation-based mechanism involves the gradient descent method while finding the rotational matrix in anisotropic scaling. The learning rate initialization in the gradient descent method involves manual intervention which makes it ineffective and unrealistic. The study finds that the incorporation of the Adaptive gradient algorithm as an optimization technique for MEQA is more suitable as it determines the learning rate adaptively from data and ensures convergence. This adaptability is advantageous for sparse datasets and enhances time complexity. The bound for convergence is independent of the initial learning rate which makes it more efficient. Experimental analysis on five benchmark synthetic datasets over five feature extraction methods demonstrates the efficacy of Adagrad in MEQA and shows improved anisotropic scaling through time and performance efficiency. Comparative analyses with alternative optimization algorithms highlight Adagrad’s favourable characteristics.