The composition of oxide glasses is characterized by high dimensionality and sparsity, making it challenging to establish high-precision predictive models. Therefore, feature extraction is essential. This study focuses on the optical properties of oxide glasses (refractive index and Abbe number), utilizing autoencoder (AE) and machine learning techniques to achieve automated feature extraction. The results indicate that compared to standalone neural networks (NN), AE-NN transforms unsupervised learning into supervised learning, reducing feature dimensions while improving model accuracy. Specifically, for the refractive index dataset, the dimensionality was reduced from 63 to 25, with a corresponding test set coefficient of determination (R2) of 0.95. For the Abbe number dataset, the dimensionality was reduced from 61 to 30, with a corresponding test set R2 of 0.97, demonstrating the effectiveness of the feature extraction method. Regarding interpretability, analyzing the encoder weight matrix of the AE-NN identified the importance of original features, with Co and Y being the most significant for both refractive index and Abbe number. Additionally, the application of the feature extraction method in machine learning models shows its generality in improving model performance, particularly for non-ensemble models such as Support Vector Regression (SVR) or k-Nearest Neighbors (KNN), exhibiting significant accuracy enhancements. Finally, targeting lanthanide glasses, the established predictive model successfully identified novel optical glasses with high refractive index and Abbe number in the La2O3-Bi2O3 and La2O3-Ta2O5 systems, presenting new possibilities for optical component design.