Spatial Models or Random Forest? Evaluating the Use of Spatially Explicit Machine Learning Methods to Predict Employment Density around New Transit Stations in Los Angeles

Abstract

The increasing use of "new" machine learning techniques, such as random forest, provides an impetus to researchers to better understand the role of space in these models. Thus, this article develops an approach for constructing spatially explicit random forest models by including spatially lagged variables to mirror various spatial econometric specifications in order to test their comparative performance against traditional spatial and nonspatial regression models for predicting block-level employment density around new transit stations in Los Angeles. This article employs a "post hoc" testing approach to isolate the impact of a particular variable (transit proximity)-and supplemental diagnostics (such as partial dependence plots and permutation importances)-to help inform explanatory relationships. The results indicate that random forest models slightly outperform spatial econometric models, and the inclusion of spatial lag parameters modestly improves random forest model accuracy-the best-fit spatial random forest model demonstrates 84.61% accuracy in predicting post-construction employment density around newly built transit stations, compared to 81.88% for the best-fit spatial econometric model and 84.37% for the nonspatial random forest model. However, given these somewhat small differences, it is not possible to conclude that the random forest approach is clearly superior to traditional spatial econometric models from these results alone.

Using the web to predict regional trade flows: data extraction, modelling, and validation

“…This paper is aligned with current epistimological debates in geography (Singleton and Arribas-Bel 2021;Credit 2021) and economics (Kleinberg et al 2015) regarding the role of machine learning algorithms in making out-of-sample predictions of data instead of focusing on explanatory research frameworks. Simply put, the above advocate towards the use of ML algorithms, such as RF, as they outperform ordinary least squares -still one of the widely used estimators to model interregional trade flows -in out-of-sample predictions even when using moderate size training datasets and limited number of predictors (Mullainathan and Spiess 2017;Athey and Imbens 2019).…”

“…Sinha et al (2019) open a dialogue on the need for spatial ensemble learning approaches, such as RF, aimed to be used with spatial data with high autocorrelation and heterogeneity. Credit (2021) introduced spatially explicit RF to predict employment density in Los Angeles. Guns and Rousseau (2014) use RF to predict and recommend high-potential research collaborations, which have not yet been materialised.…”

A House Price Modeling Based on Clustering and Kriging: The Medellín Case

“…The development of artificial intelligence techniques and Geographic Information Systems (GISs) has produced a series of studies that use methods such as Artificial Neural Networks (ANNs) (McCluskey et al, 2012;Selim, 2009;Mimis et al, 2013;Vo et al, 2015), decision tree models (McCluskey et al, 2014;Reyes-Bueno et al, 2018), and clustering (Gabrielli et al, 2017;Napoli et al, 2017). Moreover, several studies on the application of machine learning to mass appraisal have been published and obtained successful results using the random forest method (Antipov and Pokryshevskaya, 2012;Čeh et al, 2018;Credit, 2021).…”

Assessing the Impact of Ferry Transit on Urban Crime

“…These results have been obtained by using the traditional differencein-differences (DID) methodology at various radii around the stations, and supplementing it with two additional major considerations. The first important consideration is the different treatment propensities in the treated stations, examined using Causal Random Forests (CRF), a method which originates in machine learning and has been gaining popularity in both the geospatial and economic literature (Davis and Heller 2020;Deines, Wang, and Lobell 2019;Zhang et al 2018;Hoffman and Mast 2019;Credit 2021;Ho et al 2007). The second important consideration is the tendency of CRF to explore the space of interaction effects between the independent variables, which mirrors manual inclusions such as individual time fixed effects.…”

“…Analyzing the effect of a single transit line in this context is not straightforward: multiple factors can be at play that can influence crime events, including the transit services in the area of investigation. Crucially for this study, the NYC Ferry is a brand new transit option, opened in three waves in May, July, and September 2017, with numerous expansions planned along 2019-2021(NYC Ferry 2019b. We focused on weekly crime trend analysis from two years before the start of the service operation, to two years after.…”