Generates a nonlinear regression based on partial moment quadrant means.
Usage
NNS.reg(
x,
y,
factor.2.dummy = TRUE,
order = NULL,
dim.red.method = NULL,
tau = NULL,
type = NULL,
point.est = NULL,
location = "top",
return.values = TRUE,
plot = TRUE,
plot.regions = FALSE,
residual.plot = TRUE,
confidence.interval = NULL,
threshold = 0,
n.best = NULL,
smooth = FALSE,
noise.reduction = "off",
dist = NULL,
ncores = NULL,
point.only = FALSE,
multivariate.call = FALSE
)Arguments
- x
a vector, matrix or data frame of variables of numeric or factor data types.
- y
a numeric or factor vector with compatible dimensions to
x.- factor.2.dummy
logical;
TRUE(default) Automatically augments variable matrix with numerical dummy variables based on the levels of factors.- order
integer; Controls the number of partial moment quadrant means. Users are encouraged to try different
(order = ...)integer settings with(noise.reduction = "off").(order = "max")retains distinct predictor locations. A perfect training fit requires compatible responses at tied locations and interpolation at those locations; duplicate predictors with different responses cannot all be interpolated.- dim.red.method
options: ("cor", "NNS.dep", "NNS.caus", "all", "equal",
numeric vector, NULL) method for determining synthetic X* coefficients (per Dana and Dawes (2004)). Selection of a method automatically engages the dimension reduction regression. The default isNULLfor full multivariate regression.(dim.red.method = "NNS.dep")uses NNS.dep for nonlinear dependence weights, while(dim.red.method = "NNS.caus")uses NNS.caus for directional association weights.(dim.red.method = "cor")uses Spearman rank correlation for weights.(dim.red.method = "all")averages all methods for further feature engineering.(dim.red.method = "equal")uses unit weights. Alternatively, user can specify a numeric vector of coefficients. Named vectors must exactly match the encoded predictor columns and are matched by name; unnamed vectors follow encoded column order.- tau
options("ts", NULL);
NULL(default) To be used in conjunction with(dim.red.method = "NNS.caus")or(dim.red.method = "all"). If the regression is using time-series data, set(tau = "ts")to use the time-series lag convention.- type
NULL(default). To perform a classification, set to(type = "CLASS"). Like a logistic regression, it is not necessary for target variable of two classes e.g. [0, 1]: a numeric response with exactly the values 0 and 1, or a factor, character or logical response, is classified automatically. Usetype = "REG"to fit a numeric response as regression, including a two-valued response, with one or more predictors.(type = "XONLY")is the single-predictor regression alias and keeps the default x-only partition.- point.est
a numeric or factor vector with compatible dimensions to
x. Returns the fitted valuey.hatfor any value ofx. Columns are matched toxby name when the two share the same predictor names, and positionally when they share no names at all (as incbind(test.x_1, test.x_2)againstcbind(x_1, x_2)). A CSV UTF-8 byte-order mark in the first header (for example,MalebecomingX...Male) is repaired before matching names. Other partial overlaps, including misspelled predictors, are ambiguous and return an error.- location
Sets the legend location within the plot, per the
xandyco-ordinates used in base graphics legend.- return.values
logical;
TRUE(default), set toFALSEin order to only display a regression plot and call values as needed.- plot
logical;
TRUE(default) To plot regression.- plot.regions
logical;
FALSE(default). Generates 3d regions associated with each regression point for multivariate regressions. Note, adds significant time to routine.- residual.plot
logical;
TRUE(default) To ploty.hatandY.- confidence.interval
numeric [0, 1];
NULL(default) Plots the associated confidence interval with the estimate and addsconf.int.pos/conf.int.negtoFitted.xy. The intervals are global: they are built from the type-8 quantiles of the in-sample residuals (no per-segment standard errors are returned). Also applies the same level for the prediction intervals.- threshold
numeric [0, 1];
(threshold = 0)(default) Sets the threshold for dimension reduction of independent variables when(dim.red.method)is notNULL. Predictors whose coefficient magnitude is belowthresholdare dropped;NULLorNAmeans 0. For classification, a vector with one cutoff per class (named by class, or in class order) applies each cutoff to that class's index, as NNS.stack can report inNNS.dim.red.threshold.- n.best
integer;
NULL(default) Sets the number of nearest regression points to use in weighting for multivariate regression atfloor((1 - D) * sqrt(nrow(x)))(at least 1, at most the number of regression points), where D averages the copula summaries with response y and -y so the smoothing choice is unchanged by response reflection.(n.best = "all")will select and weight all generated regression points. Analogous tokin ak Nearest Neighborsalgorithm. Different values ofn.bestare tested using cross-validation in NNS.stack.NAselects the default, asNULLdoes.- smooth
logical;
FALSE(default) Applies a smoothing spline instead of local linear fit to regression points.- noise.reduction
the method of determining regression points options: ("mean", "median", "mode", "off"); In low signal:noise situations,
(noise.reduction = "mean")uses means for NNS.dep restricted partitions,(noise.reduction = "median")uses medians instead of means for NNS.dep restricted partitions, while(noise.reduction = "mode")uses modes instead of means for NNS.dep restricted partitions.(noise.reduction = "off")uses an overall central tendency measure for partitions.- dist
options:(NULL, "NNS", "L1", "L2", "FACTOR") the method of distance calculation.
dist = NULLis the default and selects the native blended NNS distance, \(sum(abs(z) + z^2)\), over range-normalized coordinates.dist = "NNS"is an explicit alias for the default.dist = "L2"selects Euclidean distance;dist = "L1"selects Manhattan distance;dist = "FACTOR"uses a frequency.- ncores
integer; value specifying the number of cores to be used in the parallelized procedure. If NULL (default), the number of cores to be used is equal to the number of cores of the machine - 1 for multivariate regression predictions large enough to repay them, and one otherwise. The estimates do not depend on the number of cores.
- point.only
Internal argument for abbreviated output.
- multivariate.call
Internal argument for multivariate regressions.
Value
UNIVARIATE REGRESSION RETURNS THE FOLLOWING VALUES:
"R2"is squared Pearson correlation between fitted and observed responses (Racine–Hastie goodness of fit), or fitted classification accuracy;"SE"returns the overall standard error of the estimate betweenyandy.hat;"Prediction.Accuracy"returns the proportion of fitted class decisions equal to the observed class;"derivative"for the coefficient of thexand its applicable range;"Point.est"for the predicted value generated;"pred.int"lower and upper prediction intervals for the"Point.est"returned using the"confidence.interval"provided;"regression.points"provides the points used in the regression equation for the given order of partitions;"Fitted.xy"returns adata.frameofx,y,y.hat,resid,NNS.ID,gradient;
MULTIVARIATE REGRESSION RETURNS THE FOLLOWING VALUES:
"R2"is squared Pearson correlation between fitted and observed responses (Racine–Hastie goodness of fit), or fitted classification accuracy;"equation"returns the numerator of the synthetic X* dimension reduction equation as adata.frameconsisting of regressor and its coefficient. Denominator is the number of non-zero coefficients (abs(coefficient) > 0), returned in last row ofequationdata.frame."x.star"returns the synthetic X* as a vector;"rhs.partitions"returns the partition points for each regressorx;"RPM"provides the Regression Point Matrix, the points for eachxused in the regression equation for the given order of partitions;"Point.est"returns the predicted value generated;"pred.int"lower and upper prediction intervals for the"Point.est"returned using the"confidence.interval"provided;"Fitted.xy"returns adata.frameof the regressors,y,y.hat,NNS.IDandresidualsfor full multivariate regression, or of the syntheticx,y,y.hat,NNS.ID,gradientandresidualsfor dimension reduction regression.confidence.intervaladdsconf.int.negandconf.int.pos.
Details
For regression, R2 measures association between fitted and observed responses: it is \(cor(y, \hat y)^2\), with the package's finite constant-vector convention. It is not \(1-SSE/SST\); a biased or incorrectly scaled prediction can have R2 = 1. Evaluate held-out prediction errors for predictive accuracy.
The default partition order combines a finite-partition dependence score with the average of the duplicated-predictor copula score under both response orientations. This complexity heuristic is invariant to reflecting the response; the underlying directional dependence and copula definitions are unchanged. A supplied order bypasses this heuristic.
Univariate regression linearly interpolates the selected regression points and extends the first or last segment beyond their range (unless a smoothing spline is requested). Endpoint slopes can be unstable with sparse or noisy boundary points. Multivariate distance weighting is a weighted average of regression-point responses, so it does not automatically continue a trend outside the training support. Neither construction supplies a general extrapolation guarantee. With the default n.best = NULL, purely categorical predictors use exact observed joint cells, and an unobserved joint combination uses its nearest cells. An explicit n.best weights that many nearest cells, as for numeric predictors. A declared factor level with no training rows is accepted in point.est and is placed by the same distance rule. A coordinate with exactly two observed values retains both terminal cells, including a dummy column in a mixed design. Other finite-order numeric coordinates retain interval cells; order = "max" uses exact joint support.
Classification forms an indicator for each observed class, retains the empirical class frequencies in each predictor cell, and interpolates or distance-weights those probability vectors. Class labels never enter numeric interpolation. Point.probabilities, Fitted.probabilities, and RPM.probabilities expose the corresponding vectors. Tied class decisions follow first appearance in the training response. Factor and character responses retain numeric output codes with class.levels; numeric responses retain their supplied values. For dimension reduction, one signed NNS index is fitted per class indicator, and the resulting probabilities are normalized across classes; equation is then a list indexed by class. confidence.interval returns Prediction.set, the smallest leading set of classes attaining the requested estimated probability mass. These are descriptive probability sets, not calibrated coverage guarantees; nominal classes have no numeric prediction interval.
Note
Please ensure
point.estis of compatible dimensions tox, error message will ensue if not compatible.Like a logistic regression, the
(type = "CLASS")setting is not necessary for target variable of two classes e.g. [0, 1]. For other numeric responses settype = "CLASS"; class values can be any distinct numbers.For low signal:noise instances, increasing the dimension may yield better results using
NNS.stack(cbind(x,x), y, method = 1, ...).
References
Viole, F. and Nawrocki, D. (2013) "Nonlinear Nonparametric Statistics: Using Partial Moments" (ISBN: 1490523995, 2nd edition: https://ovvo-financial.github.io/NNS/book/)
Vinod, H. and Viole, F. (2017) "Nonparametric Regression Using Clusters" doi:10.1007/s10614-017-9713-5
Vinod, H. and Viole, F. (2018) "Clustering and Curve Fitting by Line Segments" doi:10.20944/preprints201801.0090.v1
Viole, F. (2020) "Partitional Estimation Using Partial Moments" doi:10.2139/ssrn.3592491
Dana, J., and Dawes, R. M. (2004). The Superiority of Simple Alternatives to Regression for Social Science Predictions. Journal of Educational and Behavioral Statistics, 29(3), 317–331.
Examples
if (FALSE) { # \dontrun{
set.seed(123)
x <- rnorm(100) ; y <- rnorm(100)
NNS.reg(x, y)
## Manual {order} selection
NNS.reg(x, y, order = 2)
## Maximum {order} selection
NNS.reg(x, y, order = "max")
## For Multiple Regression:
x <- cbind(rnorm(100), rnorm(100), rnorm(100)) ; y <- rnorm(100)
NNS.reg(x, y, point.est = c(.25, .5, .75))
## For Multiple Regression based on Synthetic X* (Dimension Reduction):
x <- cbind(rnorm(100), rnorm(100), rnorm(100)) ; y <- rnorm(100)
NNS.reg(x, y, point.est = c(.25, .5, .75), dim.red.method = "cor", ncores = 1)
## IRIS dataset examples:
# Dimension Reduction:
NNS.reg(iris[,1:4], iris[,5], dim.red.method = "cor", order = 5, ncores = 1)
# Dimension Reduction using causal weights:
NNS.reg(iris[,1:4], iris[,5], dim.red.method = "NNS.caus", order = 5, ncores = 1)
# Multiple Regression:
NNS.reg(iris[,1:4], iris[,5], order = 2, noise.reduction = "off")
# Classification:
NNS.reg(iris[,1:4], iris[,5], point.est = iris[1:10, 1:4], type = "CLASS")$Point.est
## To call fitted values:
x <- rnorm(100) ; y <- rnorm(100)
NNS.reg(x, y)$Fitted
## To call partial derivative (univariate regression only):
NNS.reg(x, y)$derivative
} # }