<p>The effective management of modern Digital Service Networks (DSNs) and the application of Large Language Models (LLMs) for their lifecycle automation are contingent upon the availability of complete, high-fidelity operational data. However, telemetry from complex infrastructures such as cloud services and IoT ecosystems is frequently plagued by missing values, which severely degrades the performance of downstream analysis and control tasks. While data imputation has been extensively studied, existing methods often fail to leverage sparse but critical supervisory signals, such as labeled states or anomaly flags, which are crucial in DSN contexts. To address this foundational challenge, in this paper, we propose a novel semi-supervised generative adversarial network model, named SEGAN, for missing value imputation. It consists of three modules: generator, discriminator, and classifier. In the SEGAN model, the classifier enables the generator to make full use of the known data and label information when predicting missing data values. In addition, the SEGAN model introduces a missing hint matrix to allow the discriminator to distinguish between the known data and data filled by the generator more effectively. This study theoretically proves that the SEGAN model, which introduces a classifier and a missing hint matrix, can learn the real known data distribution characteristics when reaching Nash equilibrium. Extensive experiments on three public real-world datasets demonstrated that, SEGAN yields an accuracy gain of more than 10%, compared with state-of-the-art approaches.</p>