A Case for Exploration: Exploratory Data Analysis in Neural Networks for Renal Tumor Classification

Exploratory data analysis provides insights into the structure and representation of data. By understanding data, more representative networks can be created without total reliance on implicit processes to match internal representations with expectation. The aim of
this paper is to demonstrate how the usage of exploratory data analysis combined with knowledge from subject matter experts can increase
the performance of networks and reduce reliance on mechanics. We use
manual transformations based on data analysis to represent our data in
a way that resembles the abstract processes that experts use to classify
renal tumors. Compared to a baseline network, we achieved a significant increase in accuracy, 93.3% to 98.64%, without loss in external
validations, while reducing the network size considerably. The reliance
on internal mechanics to achieve the expected representation is flawed
and exploratory data analysis can mitigate some of the pitfalls of this
approach, leading to increased performance and a reduction in required
compute power.