--- title: "Random Forest Regression with ggRandomForests" author: "John Ehrlinger" date: today format: html: fig-format: png fig-dpi: 96 toc: true toc-depth: 3 html-math-method: mathjax code-fold: false bibliography: ggRandomForests.bib vignette: > %\VignetteIndexEntry{Random Forest Regression with ggRandomForests} %\VignetteEngine{quarto::html} %\VignetteEncoding{UTF-8} --- ::: {.callout-warning} ## Work in progress This vignette is under active development. Code examples and narrative may change before the next release. ::: ```{r setup, include=FALSE} # vignette dir under R CMD check, package root under an interactive knit .fo <- c("_fig_optim.R", file.path("vignettes", "_fig_optim.R")) if (any(file.exists(.fo))) source(.fo[file.exists(.fo)][1]) knitr::opts_chunk$set( fig.align = "center", fig.width = 7, fig.height = 5, message = FALSE, warning = FALSE, comment = "#>" ) options(mc.cores = 1, rf.cores = 1) ``` # Introduction Random forests [@Breiman:2001] are a non-parametric ensemble method that requires no distributional or functional assumptions on how covariates relate to the response. The method builds a large collection of de-correlated decision trees via bootstrap aggregation (bagging) and random feature selection, then averages their predictions to smooth out the noise any single tree carries. The **randomForestSRC** package [@Ishwaran:RFSRC:2014] provides a unified implementation for survival, regression, and classification forests. **ggRandomForests** extracts tidy data objects from `rfsrc` fits and renders them with **ggplot2** [@Wickham:2009], making it straightforward to explore how a forest is constructed, which variables matter, and how the response depends on individual predictors. This vignette demonstrates a complete random forest regression workflow on the Boston Housing data set [@Harrison:1978; @Belsley:1980]: 1. **Data exploration** --- EDA scatter panels, variable descriptions 2. **Growing the forest** --- fitting an RF, checking OOB error convergence 3. **Variable selection** --- VIMP and minimal depth via `max.subtree()` 4. **Dependence plots** --- variable dependence and partial dependence via `gg_variable()` and `gg_partial_rfsrc()` 5. **Variable interactions** --- conditioning plots and partial dependence surfaces ```{r packages} library(ggplot2) library(dplyr) library(tidyr) library(randomForestSRC) if (requireNamespace("ggRandomForests", quietly = TRUE)) { library(ggRandomForests) } else if (requireNamespace("pkgload", quietly = TRUE)) { pkgload::load_all(export_all = FALSE, helpers = FALSE, attach_testthat = FALSE) } else { stop("Install ggRandomForests (or pkgload for dev builds) to render this vignette.") } theme_set(theme_bw()) ``` # Data: Boston Housing Values The Boston Housing data [@Harrison:1978; @Belsley:1980] is a standard benchmark for regression. It contains data for 506 census tracts of Boston from the 1970 census. The objective is to predict the median value of owner-occupied homes (`medv`, in \$1000s) from 13 predictors covering crime, zoning, industry, environmental quality, and housing characteristics. We use the copy from the **MASS** package [@mass:2002]. ```{r data-load} data(Boston, package = "MASS") Boston$chas <- as.logical(Boston$chas) # nolint: object_name_linter ``` ```{r variable-labels} st_labs <- c( crim = "Crime rate by town", zn = "Residential land zoned > 25k sq ft (%)", indus = "Non-retail business acres (%)", chas = "Borders Charles River", nox = "Nitrogen oxides (10 ppm)", rm = "Rooms per dwelling", age = "Units built before 1940 (%)", dis = "Distance to employment centers", rad = "Highway accessibility index", tax = "Property tax rate per $10,000", ptratio = "Pupil-teacher ratio", black = "Proportion of Black residents", lstat = "Lower status population (%)", medv = "Median home value ($1000s)" ) ``` ## Exploratory data analysis We plot each predictor against the response (`medv`), coloring by the sole categorical variable (`chas`, whether the tract borders the Charles River). A loess smooth highlights the marginal trend. ```{r eda, fig.cap="EDA: each predictor vs. median home value, colored by Charles River indicator."} dta <- Boston |> pivot_longer(c(-medv, -chas), names_to = "variable", values_to = "value") ggplot(dta, aes(x = medv, y = value, color = chas)) + geom_point(alpha = 0.4) + geom_smooth(aes(x = medv, y = value), color = "gray30", inherit.aes = FALSE, se = FALSE) + labs(y = "", x = st_labs["medv"]) + scale_color_brewer(palette = "Set2") + facet_wrap(~variable, scales = "free_y", ncol = 3) ``` Even from this simple view, two relationships stand out: `medv` against `lstat` (lower status %) and `medv` against `rm` (rooms per dwelling). Keep those two in mind --- we expect the random forest to rank them as the most important predictors, and the rest of the vignette comes back to check. # Growing a Random Forest We grow a regression forest using all 13 predictors. The `rfsrc()` function detects the regression family from the continuous response. ```{r rfsrc-fit} rfsrc_Boston <- rfsrc(medv ~ ., data = Boston, # nolint: object_name_linter ntree = 100, importance = TRUE, err.block = 5) rfsrc_Boston ``` The forest grew `r rfsrc_Boston$ntree` trees, splitting on `r rfsrc_Boston$mtry` randomly selected candidate variables at each node, and stopping at a minimum terminal node size of `r rfsrc_Boston$nodesize`. ## OOB error convergence ```{r error-plot, fig.height=4, fig.cap="OOB mean squared error vs. number of trees."} gg_e <- gg_error(rfsrc_Boston) gg_e <- gg_e |> filter(!is.na(error)) class(gg_e) <- c("gg_error", class(gg_e)) plot(gg_e) ``` The error stabilizes well before 500 trees, indicating the forest is large enough for reliable predictions. ## OOB predictions ```{r rfsrc-predicted, fig.height=4, fig.cap="OOB predicted median home values. Points are jittered; boxplot shows the distribution."} plot(gg_rfsrc(rfsrc_Boston), alpha = 0.5) + coord_cartesian(ylim = c(5, 49)) ``` Each point is a single tract's OOB prediction. The distribution is a sanity check --- we are more interested in the *why* behind these predictions. # Variable Selection ## Variable importance (VIMP) VIMP, computed by [randomForestSRC](https://www.randomforestsrc.org), measures the increase in OOB prediction error when a variable's values are randomly permuted across the out-of-bag observations [@Breiman:2001]. Permutation severs the variable's link to the response on purpose: if breaking that link hurts accuracy, the variable is carrying real signal. Large positive values mean the variable is essential; negative values suggest it is no more informative than noise. ```{r vimp-plot, fig.cap="VIMP ranking. Longer blue bars indicate more important variables."} plot(gg_vimp(rfsrc_Boston), lbls = st_labs) ``` `lstat` and `rm` dominate, with a clear gap to the remaining predictors. All VIMP values are positive, indicating every predictor contributes at least marginally. The permutation approach contrasts with varPro release-rule importance [@Lu2024varpro], available through `gg_varpro()`. Rather than perturbing data synthetically, varPro compares local estimators on the observed data directly: no permutation, no manufactured feature values. Because the two methods measure fundamentally different things, a variable can rank high under one and low under the other. When they agree, the evidence is strong; when they disagree, that disagreement itself is worth investigating, pointing either to a variable whose effect is highly non-linear or to one that matters only in combination with others. ## Minimal depth Minimal depth [@Ishwaran:2010] ranks variables by how close to the root node they first split, on average. Variables that partition large portions of the population early are considered most important. ```{r varsel} md_Boston <- max.subtree(rfsrc_Boston) # nolint: object_name_linter ``` The threshold is `r round(md_Boston$threshold, 2)`, selecting `r length(md_Boston$topvars)` variables: `r paste(md_Boston$topvars, collapse = ", ")`. Both VIMP and minimal depth agree on the dominance of `lstat` and `rm`. We use the minimal depth top variables for the remainder of the analysis. ```{r select-vars} xvar <- md_Boston$topvars ``` # SHAP Analysis VIMP and varPro both rank *how much* a variable matters, averaged over the whole forest. They cannot tell you how much a variable mattered for one specific tract's prediction --- that requires a different kind of accounting. SHAP (SHapley Additive exPlanations) borrows an idea from cooperative game theory: treat the 13 predictors as players splitting a payout, and ask how much each one contributed to this tract's predicted `medv`, averaged fairly over every order the players could have joined the game. The result is a signed contribution per predictor per tract, and those contributions add up exactly: the forest's overall average prediction (the *baseline*) plus the sum of one tract's contributions equals that tract's actual predicted value. No other importance measure in this vignette makes that promise. `gg_shap()` computes these contributions by calling [kernelshap](https://cran.r-project.org/package=kernelshap), which needs one prediction per predictor per background draw, per tract explained. Scoring all 506 tracts is expensive, so we explain a random sample of 25 instead, plenty to see the patterns at a fraction of the cost. ```{r shap-fit} set.seed(42) shap_sample <- Boston[sample(nrow(Boston), 25), setdiff(colnames(Boston), "medv")] gg_shp <- gg_shap(rfsrc_Boston, newdata = shap_sample, bg_n = 30) ``` ## SHAP importance Averaging the absolute contribution of each variable across the 25 tracts gives a ranking directly comparable to VIMP. ```{r shap-importance-plot, fig.cap="Mean absolute SHAP value per predictor."} plot(gg_shp, type = "importance") ``` `lstat` and `rm` come out on top again, same as VIMP and minimal depth --- three different mechanisms, one answer. That agreement is reassuring, but it's the next two plots where SHAP earns its keep. ## SHAP beeswarm The beeswarm plot is one dot per tract per variable: its horizontal position is the SHAP contribution (negative pulls the prediction down, positive pushes it up), and its color is that tract's value for the variable, scaled low to high within each row so the pattern doesn't get washed out by variables on very different scales. ```{r shap-beeswarm-plot, fig.cap="SHAP beeswarm: every dot is one tract's contribution for one predictor."} plot(gg_shp, type = "beeswarm") ``` `lstat` shows a clean color gradient: the yellow (high lower-status population) dots sit on the negative side, the purple (low) dots on the positive side. `rm` runs the other way. Both track what the EDA scatterplot already hinted at, but now stated as a contribution to an individual prediction rather than a marginal trend. ## SHAP dependence Plotting one variable's SHAP contribution against its own value is the closest SHAP gets to a partial dependence plot, tract by tract instead of averaged. ```{r shap-dependence-plot, fig.cap="SHAP dependence for lstat."} plot(gg_shp, type = "dependence", xvar = "lstat") ``` The downward slope matches the sign of the correlation directly: tracts with more lower-status population get pulled below the baseline prediction, tracts with less get pushed above it. The next section asks the same question a second way, partial dependence averaged over the whole forest instead of tract by tract, and the two views should agree on the shape even though they're built from entirely different mechanics. # Variable Dependence ## Variable dependence plots Variable dependence shows each tract's OOB predicted `medv` plotted against a predictor, with a loess smooth indicating the trend. ```{r vardep-panel, fig.cap="Variable dependence for top predictors (minimal depth rank order)."} gg_v <- gg_variable(rfsrc_Boston) plot(gg_v, xvar = xvar, panel = TRUE, alpha = 0.5) + labs(y = st_labs["medv"], x = "") ``` The panels confirm what EDA suggested: `medv` decreases sharply with `lstat` and increases with `rm`, both in strongly non-linear ways. The remaining variables show weaker but still discernible trends. ```{r vardep-chas, fig.height=4, fig.cap="Variable dependence for Charles River (categorical)."} plot(gg_v, xvar = "chas", alpha = 0.4) + labs(y = st_labs["medv"]) ``` Most tracts do not border the Charles River, and the predicted value distributions largely overlap --- consistent with `chas` ranking last in both VIMP and minimal depth. ## Partial dependence Partial dependence integrates out the effects of all other covariates, giving a risk-adjusted view of each predictor's marginal effect [@Friedman:2000]: $$ \tilde{f}(x) = \frac{1}{n} \sum_{i=1}^n \hat{f}(x, x_{i,o}) $$ We use `gg_partial_rfsrc()`, which calls `randomForestSRC::partial.rfsrc()` directly and returns a `gg_partial_rfsrc` object. The quickest path to a figure is `plot(pd)`, which sorts out continuous and categorical variables for you. When you want to control the layout yourself, the underlying data is in `pd$continuous`. ```{r partial-dep, fig.cap="Partial dependence for top predictors."} pd <- gg_partial_rfsrc(rfsrc_Boston, xvar.names = xvar) # Quick S3 plot — works out of the box for the standard regression case plot(pd) ``` For a publication-ready layout with custom axis labels, access the underlying data frame directly: ```{r partial-dep-custom, fig.cap="Partial dependence (custom styling)."} ggplot(pd$continuous, aes(x = x, y = yhat)) + geom_line(color = "steelblue", linewidth = 1) + facet_wrap(~name, scales = "free_x") + labs(y = st_labs["medv"], x = "") + theme_bw() ``` `lstat` shows a strongly concave relationship, while `rm` stays flat below about 6 rooms and then climbs sharply. Shapes like these are awkward to capture with a simple parametric transform, since you would have to guess the form in advance, but the random forest picks them up on its own. # Variable Interactions and Conditioning Plots ## Conditioning on a categorical variable The simplest coplot conditions on a categorical variable. Here we examine `medv` vs. `lstat`, split by Charles River status: ```{r coplot-chas, fig.height=4, fig.cap="Variable dependence on lstat, conditional on Charles River."} gg_v$chas_label <- ifelse(gg_v$chas, "Borders Charles River", "Does not border") plot(gg_v, xvar = "lstat", alpha = 0.5) + labs(y = st_labs["medv"], x = st_labs["lstat"]) + theme(legend.position = "none") + facet_wrap(~chas_label) ``` The decreasing trend holds in both groups, with slightly higher values along the Charles River at every `lstat` level. ## Conditioning on a continuous variable To investigate the `lstat`--`rm` interaction, we bin `rm` into six quantile groups using `quantile_pts()` and facet: ```{r coplot-rm, fig.cap="Predicted medv vs. lstat, conditional on rooms-per-dwelling groups."} rm_pts <- quantile_pts(rfsrc_Boston$xvar$rm, groups = 6, intervals = TRUE) gg_v$rm_grp <- cut(rfsrc_Boston$xvar$rm, breaks = rm_pts) levels(gg_v$rm_grp) <- paste("rm in", levels(gg_v$rm_grp)) plot(gg_v, xvar = "lstat", alpha = 0.5) + labs(y = st_labs["medv"], x = st_labs["lstat"]) + theme(legend.position = "none") + scale_color_brewer(palette = "Set3") + facet_wrap(~rm_grp) ``` Median values decrease with `lstat` within every `rm` group, but the intercept shifts upward with more rooms. Smaller homes in low-`lstat` (high-status) neighborhoods still command high prices. The complement view --- `medv` vs. `rm`, conditional on `lstat` groups --- completes the picture: ```{r coplot-lstat, fig.cap="Predicted medv vs. rooms, conditional on lower-status groups."} lstat_pts <- quantile_pts(rfsrc_Boston$xvar$lstat, groups = 6, intervals = TRUE) gg_v$lstat_grp <- cut(rfsrc_Boston$xvar$lstat, breaks = lstat_pts) levels(gg_v$lstat_grp) <- paste("lstat in", levels(gg_v$lstat_grp)) plot(gg_v, xvar = "rm", alpha = 0.5) + labs(y = st_labs["medv"], x = st_labs["rm"]) + theme(legend.position = "none") + scale_color_brewer(palette = "Set3") + facet_wrap(~lstat_grp) ``` The `rm` effect is strongest in low-`lstat` tracts (bottom-left panels) and nearly flat in high-`lstat` tracts, confirming a meaningful interaction. # Partial Dependence Surface To visualize the joint partial dependence of `medv` on `lstat` and `rm`, we compute partial dependence on a grid: 6 values of `rm`, each evaluated at 25 points along `lstat`. ```{r surface-data, cache=TRUE} rm_grid <- quantile_pts(rfsrc_Boston$xvar$rm, groups = 6) surface_list <- lapply(rm_grid, function(rm_val) { newx <- rfsrc_Boston$xvar newx$rm <- rm_val pd_rm <- gg_partial_rfsrc(rfsrc_Boston, xvar.names = "lstat", newx = newx) df <- pd_rm$continuous df$rm <- rm_val df }) surface_df <- bind_rows(surface_list) ``` ```{r pd-surface, fig.cap="Partial dependence surface: median home value as a function of lstat and rm. Fill color is the predicted median value."} ggplot(surface_df, aes(x = x, y = rm, fill = yhat)) + geom_tile() + scale_fill_viridis_c(name = "Median Value\n($1000s)") + labs(x = "Lower Status (%)", y = "Rooms per Dwelling") + theme_bw() ``` The surface confirms the strong interaction: home values are highest when `lstat` is low and `rm` is high (upper-left corner), dropping steeply along both axes. The non-planar curvature --- particularly the sharp step near `rm` = 7 --- demonstrates the kind of complex, non-linear structure that random forests capture naturally. # Conclusion We have walked a full random forest regression analysis with **randomForestSRC** and **ggRandomForests**, and the pieces line up: - `gg_error()` showed the OOB error settling well before 500 trees. - VIMP (`gg_vimp()`) and minimal depth (`max.subtree()`) agreed on the same story: `lstat` and `rm` dominate, with a clear gap to everything else. - `gg_variable()` traced strongly non-linear predictor--response curves, the same shapes the raw-data EDA hinted at. - Partial dependence from `gg_partial_rfsrc()` gave the risk-adjusted version of those curves: concave for `lstat`, threshold-like for `rm`. - Conditioning plots and the partial dependence surface pulled out the `lstat`--`rm` interaction, with the room-size effect strongest in high-status tracts. Notice the pattern in all of this. Each `gg_*()` function returns a tidy object (often a data frame, sometimes a small list of data frames); the plotting is a separate step. Use the package's `plot()` methods when the default figure is what you want, and reach for `ggplot2` directly when it is not. # References