Application of Random Forest and SHAP Tree Explainer in Exploring Spatial (In)Justice to Aid Urban Planning

Abstract

In light of recent local, national, and global events, spatial justice provides a potentially powerful lens by which to explore a multitude of spatial inequalities. For more than two decades, scholars have been espousing the power of this concept to help develop more equitable and just communities. However, defining spatial justice and developing a methodology for quantitatively analyzing it is complicated and no agreed upon metric for examining spatial justice has been developed. Instead, individual measures of spatial injustices have been studied. One such individual measure is economic mobility. Recent research on economic mobility has revealed the importance of local geography on upward mobility and may serve as an important keystone in developing a metric for multiple place-based issues of spatial inequality. This paper seeks to explore place-based variables within individual census tracts in an effort to understand their impact on economic mobility and potentially spatial justice. The methodology relies on machine learning techniques and the results show that the best performing model is able to predict economic mobility of a census tract based on its spatial variables with 86% accuracy. The availability and density of jobs, compactness of the area, and the presence of medical facilities and underground storage tanks have the greatest influence, whereas some of the influential features are positively while the others are negatively associated. In the end, this research will allow for comparative analysis between differing geographies and also identify leading variables in the overall quest for spatial justice.

Using machine learning to identify patient characteristics to predict mortality of in-patients with COVID-19 in South Florida

“…Correlation analysis was completed only for exploratory data analysis. However, this research did not use correlation as a guideline for selecting features, as two correlated features can further improve the model accuracy when they are part of the same data set ( 32 ).…”

“…A method that performs over-sampling is the S ynthetic M inority O ver-sampling T echnique (SMOTE), by synthesizing new examples as opposed to duplicating examples ( 34 ). The SMOTE was applied to the training data set in the cross-validation to avoid the possibility of over-fitting; however, this technique was not applied to the test data set for model evaluation that prevents data leakage ( 32 ).…”

“…The color corresponded to the value of the function. As discussed before, the "red" color depicted a higher SHAP value of a feature that fell within the right distribution; on the other hand, "blue" color mapped a lower SHAP value of a feature that fell within the left distribution of the reference point "0" (32). For binary categorical variables (e.g., sex, diarrhea, diabetes, etc.…”

Explanations of Machine Learning Models in Repeated Nested Cross-Validation: An Application in Age Prediction Using Brain Complexity Features

“…However, the SHAP framework has been proposed for hold-out strategies [26][27][28][29][30], where SHAP values are computed only when a final model is trained. Some authors adopted the SHAP method with CV strategies [31][32][33][34][35][36], but SHAP was used only after the CV procedure on often unclear portions of the dataset. Recently, two related works [11,37] adopted an average of SHAP values, performing multiple trainings of the model with different undersampling of the training data and computing SHAP values on the test sets.…”

Where Are Potential Areas for Transit-Oriented Development (TOD)—Exploring the Demands for Built Environment for TOD Planning

“…Firstly, whether a specific area has a typical TOD feature was evaluated by random forest (RF). RF exhibits excellent performance in identifying urban functional areas and urban land use type, and has been adopted in a number of urban computer studies [19][20][21]. Endowed with a robust classification function and a superior algorithm (multiple decision trees randomly vote to comprehensively define the final result), RF is more prepared to avoid issues such as variable endogeneity and model overfitting [22].…”