Reproduce “Tab2Visual: Overcoming Limited Data in Tabular Data Classification Using Deep Learning with Visual Representations” on OpenResearch

About this paper

Ahmed Mamdouh, Moumen El-Melegy, Samia Ali, Ron Kikinis

Tab2Visual introduces a method for tabular data classification that converts tabular inputs into visual representations, enabling the use of deep learning models like CNNs and ViTs. This approach, which includes tailored image augmentation and transfer learning, consistently improved performance on small datasets (up to 2.1% AUC gain with augmentation) and effectively addressed challenges of data scarcity in domains like prostate cancer prediction.

Key insights

• Transforming tabular data into visual formats enables powerful deep learning vision models to be applied effectively, even with limited feature sets.
• Tailored image augmentation methods can significantly boost the performance and generalization of deep learning on small tabular datasets, yielding up to a 2.1% AUC gain.
• Transfer learning from pre-trained vision models offers substantial benefits for tabular classification, especially on small datasets, improving AUC by up to 7.5%.

Summary by alphaXiv.