Segmented R2: A Proposed Variant for Unbalanced and Skewed Dataset
Wai Keung Wong · 2024
R2(R-Squared) is an important metric that is often used for evaluating the Goodness-of-fit (GOF) or evaluating variables corelation. As a measurement of GOF, the metric is only accurate when the distribution of the observed target is equal. This research proposes a modified R2 based on the distribution interval. The proposed approach includes modification by separating the distribution range into smaller sections and evaluating sections R2. Subsequently, the section R2 can be evaluated to reflect a more realistic fit. A data set with with skewed distribution as used and prediction with Multi linear regression in which GOF is evaluated using both R2. The standard R2 and the modified R2 was implemented as GOF. This report demonstrates that, despite good GOF using conventional R2, it is actually not a well-fitting model. It is better represented using the modified R2 metrics. In short, the proposed approach indicates a preliminary introduction to a simple and effective modification to existing R2 that is ideal for representing goodness of fit in view of a non-balanced distribution of data. When compared to the nearest solutions i.e pseudo R2, the measurement is less sensitive to sample size.