Graph Neural Networks (GNNs) are among the most powerful neural network architectures currently available. Unlike traditional neural networks that solely rely on the feature vector of the data to be processed as input, GNNs utilize both the graph that represents the relationships between data points and the feature matrix of the data to optimize their feature representation. This unique capability allows GNNs to achieve superior performance in various tasks. However, GNNs are highly susceptible to noise, which can significantly degrade their performance in common tasks such as classification and prediction. This paper proposes a novel approach leveraging spectral analysis to identify and select the points most sensitive to noise for constructing the training set. By training GNNs on these noise-sensitive points, experimental results demonstrate a substantial improvement in training efficiency compared to using randomly selected training sets.