Studying the Evolution of TensorFlow Questions on Stack Overflow
Gbolahan Michael Adebesin, Sangeeta Sangeeta · Journal of Software Evolution and Process · 2026
ABSTRACT Deep learning has revolutionized various fields, including computer vision, natural language processing, and robotics. Python, with its simplicity and extensive libraries, has emerged as one of the primary programming languages for implementing deep learning models, and several frameworks like TensorFlow, Gaffe, and PyTorch are proposed by machine learning communities/companies for Python. TensorFlow is one of the most popular frameworks in Python for deep learning; hence, it is important to investigate the kind of issue that TensorFlow developers face. In this work, we analyzed 105,437 TensorFlow questions from the Stack Overflow website and answered five research questions. Our analysis reveals that, despite a decreasing trend in the number of questions asked, there is a high level of user engagement and satisfaction in the TensorFlow community. Our tag analysis reveals that “python,” “Keras,” “deep‐learning,” “machine‐learning,” and “neural‐network” are the top tags associated with TensorFlow questions. We perform topic analysis and reveal four main topics: machine learning and data processing , neural network and architecture , TensorFlow installation and usage , and TensorFlow manipulation and operation . Further analysis reveals that error , model training and prediction , optimization , data loading and preprocessing , and installation and deployment are among the most frequently occurring subtopics. ValueError is the most common error and the model training and prediction subtopic is the most affected by errors. Our analysis reveals that the TensorFlow installation and usage is the most popular and difficult topic. Our study also reveals a high percentage (i.e., 31%) of broken links in TensorFlow questions.