This research investigates the detection of pleasant fragrances using a comprehensive approach leveraging advanced machine learning techniques. Initially, a dataset comprising pleasant (Indian Jasmine) and unpleasant (Human Urine) samples, collected via GC-MS dataset is utilized. To address the challenge of missing data, a Missing Completely at Random (MCaR) based preprocessing method is applied to the dataset. Following preprocessing, a Heterogeneous Graph Convolution Network with Cell Attention (HGCNCA) is employed to accurately classify the jasmine fragrance. To further enhance the performance of the HGCNCA, a skill optimization algorithm is incorporated. This combination of preprocessing and advanced graph-based deep learning techniques aims to improve the reliability and precision of fragrance detection, providing a robust framework for distinguishing between pleasant and unpleasant odors based on analytical data. The proposed method attains higher accuracy as 99.9%.