Improving Prediction Accuracy by Controlling of Unobserved Variables via Incorporating Weakly Correlated Variables Using Unsupervised Random Forest

Title: Improving Prediction Accuracy by Controlling of Unobserved Variables via Incorporating Weakly Correlated Variables Using Unsupervised Random Forest

Authors: Tao Yang^1*, Hong-Wen Deng¹, Tianhua Niu¹

¹Center for Bioinformatics and Genomics, Department of Biostatistics and Bioinformatics, Tulane University School of Public Health and Tropical Medicine, New Orleans, LA 70112, USA
^*Corresponding Author

Abstract
Risk predictions are increasely being applied in clinical reasoning and decision making in modern medicine and epidemiological research. However, risk prediction is inevitably affected by a myriad of factors, which could affect the accuracy of the resulting prediction model. A particular challenge is to control for the effects of unobserved variables (UVs). To overcome this problem, we have implemented a novel machine learning technique known as unsupervised random forest (RF). Under a range of parameter settings, we performed extensive Monte Carlo simulations mimicking various scenarios of real-world data to assess whether an RF ensemble classifier that incorporates a set of weakly correlated variables (WCVs) could improve the accuracy of the accuracy of the prediction model in the presence of UVs. Our simulation results demonstrated that in almost all of simulation scenarios, accounting for effects of UVs has dramatically improved the overall prediction accuracy. Specifically, we found that the magnitude of such improvement depends on the number of WCVs, the correlations among WCVs, , and the non-linearity/non-additivity relationships between WCVs and UVs. Out study suggests that unsupervised RFs could have profound applications in constructing an accurate prediction model for handling UVs in clinical and epidemiological studies.