X-Git-Url: https://git.auder.net/?p=talweg.git;a=blobdiff_plain;f=pkg%2FR%2Fplot.R;h=0f895bd16483795302623b53ea983f3ab9c2cd92;hp=b720e9a16fe588f492cb6a77af54516115b94bda;hb=HEAD;hpb=469529710f56c790ae932b45d13fed2e34bcabf2

diff --git a/pkg/R/plot.R b/pkg/R/plot.R
index b720e9a..0f895bd 100644
--- a/pkg/R/plot.R
+++ b/pkg/R/plot.R
@@ -1,100 +1,102 @@
-#' @title plot curves
+#' Plot curves
 #'
-#' @description Plot a range of curves in data
+#' Plot a range of curves in data.
 #'
-#' @param data Object of class Data
+#' @inheritParams computeError
 #' @param indices Range of indices (integers or dates)
 #'
 #' @export
-plotCurves <- function(data, indices)
+plotCurves <- function(data, indices=seq_len(data$getSize()))
 {
-	yrange = range( sapply( indices, function(i) {
-		serie = c(data$getCenteredSerie(i))
-		if (!all(is.na(serie)))
-			range(serie, na.rm=TRUE)
-		c()
-	}) )
+	series = data$getSeries(indices)
+	yrange = quantile(series, probs=c(0.025,0.975), na.rm=TRUE)
 	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
-	for (i in seq_along(indices))
-	{
-		plot(data$getSerie(indices[i]), type="l", ylim=yrange,
-			xlab=ifelse(i==1,"Temps (en heures)",""), ylab=ifelse(i==1,"PM10",""))
-		if (ii < length(indices))
-			par(new=TRUE)
-	}
+	matplot(series, type="l", ylim=yrange, xlab="Time (hours)", ylab="PM10")
 }
 
-#' @title plot measured / predicted
+#' Plot error
 #'
-#' @description Plot measured curve (in black) and predicted curve (in red)
+#' Draw error graphs, potentially from several runs of \code{computeForecast()}.
+#'
+#' @param err Error as returned by \code{computeError()}
+#' @param cols Colors for each error (default: 1,2,3,...)
+#' @param agg Aggregation level ("day", "week" or "month")
 #'
-#' @param data Object return by \code{getData}
-#' @param pred Object as returned by \code{getForecast}
-#' @param index Index in forecasts
+#' @seealso \code{\link{plotCurves}}, \code{\link{plotPredReal}},
+#'   \code{\link{plotSimils}}, \code{\link{plotFbox}}, \code{\link{computeFilaments}},
+#'   \code{\link{plotFilamentsBox}}, \code{\link{plotRelVar}}
 #'
 #' @export
-plotPredReal <- function(data, pred, index)
+plotError <- function(err, cols=seq_along(err), agg="day")
 {
-	horizon = length(pred$getSerie(1))
-	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5, lwd=3)
-	measure = data$getSerie(pred$getIndexInData(index)+1)[1:horizon]
-	yrange = range( pred$getSerie(index), measure )
-	plot(measure, type="l", ylim=yrange, xlab="Temps (en heures)", ylab="PM10")
-	par(new=TRUE)
-	plot(pred$getSerie(index), type="l", col="#0000FF", ylim=yrange, xlab="", ylab="")
+	if (!is.null(err$abs))
+		err = list(err)
+	par(mfrow=c(2,2), mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
+	L = length(err)
+
+	yrange = range( sapply(1:L, function(i) err[[i]]$abs$day), na.rm=TRUE )
+	matplot(sapply( seq_len(L), function(i) err[[i]]$abs$day ), type="l",
+		xlab="Time (hours)", ylab="Mean |y - y_hat|", ylim=yrange, col=cols, lwd=2, lty=1)
+
+	agg_curves <- sapply( seq_len(L), function(i) {
+		curve <- err[[i]]$abs$indices
+		delta <- if (agg=="day") 1 else if (agg=="week") 7 else if (agg=="month") 30
+		vapply( seq(1,length(curve),delta), function(i) {
+			mean(curve[i:(i+delta-1)], na.rm=TRUE)
+		}, vector("double",1), USE.NAMES=FALSE )
+	})
+	yrange = range(agg_curves, na.rm=TRUE)
+	matplot(agg_curves, type="l", xlab=paste("Time (",agg,"s)", sep=""),
+		ylab="Mean |y - y_hat|", ylim=yrange, col=cols, lwd=2, lty=1)
+
+	yrange = range( sapply(1:L, function(i) err[[i]]$MAPE$day), na.rm=TRUE )
+	matplot(sapply( seq_len(L), function(i) err[[i]]$MAPE$day ), type="l",
+		xlab="Time (hours)", ylab="Mean MAPE", ylim=yrange, col=cols, lwd=2, lty=1)
+
+	agg_curves <- sapply( seq_len(L), function(i) {
+		curve <- err[[i]]$MAPE$indices
+		delta <- if (agg=="day") 1 else if (agg=="week") 7 else if (agg=="month") 30
+		vapply( seq(1,length(curve),delta), function(i) {
+			mean(curve[i:(i+delta-1)], na.rm=TRUE)
+		}, vector("double",1), USE.NAMES=FALSE )
+	})
+	yrange = range(agg_curves, na.rm=TRUE)
+	matplot(agg_curves, type="l", xlab=paste("Time (",agg,"s)", sep=""),
+		ylab="Mean MAPE", ylim=yrange, col=cols, lwd=2, lty=1)
 }
 
-#' @title Plot filaments
+#' Plot measured / predicted
 #'
-#' @description Plot similar days in the past + "past tomorrow", as black as distances are small
+#' Plot measured curve (in black) and predicted curve (in blue).
 #'
-#' @param data Object as returned by \code{getData}
-#' @param index Index in data
-#' @param limit Number of neighbors to consider
+#' @inheritParams computeError
+#' @param index Index in forecasts (integer or date)
 #'
 #' @export
-plotFilaments <- function(data, index, limit=60)
+plotPredReal <- function(data, pred, index)
 {
-	index = dateIndexToInteger(index, data)
-	ref_serie = data$getCenteredSerie(index)
-	if (any(is.na(ref_serie)))
-		stop("plotFilaments requires a serie without NAs")
-	L = length(ref_serie)
-	first_day = ifelse(length(data$getCenteredSerie(1)<L), 2, 1)
-	distances = sapply(first_day:(index-1), function(i) {
-		sqrt( sum( (ref_serie - data$getCenteredSerie(i))^2 ) / L )
-	})
-	# HACK to suppress NA effect while keeping indexation
-	distances[is.na(distances)] = max(distances,na.rm=TRUE) + 1
-	indices = sort(distances, index.return=TRUE)$ix[1:min(limit,index-first_day)]
-	yrange = range( ref_serie, sapply( indices, function(i) {
-		index = i - first_day + 1
-		serie = c(data$getCenteredSerie(index), data$getCenteredSerie(index+1))
-		if (!all(is.na(serie)))
-			return (range(serie, na.rm=TRUE))
-		c()
-	}) )
-	grays = gray.colors(20, 0.1, 0.9) #TODO: 20 == magic number
-	colors = c(
-		grays[ floor( 20.5 * distances[indices] / (1+max(distances[indices])) ) ], "#FF0000")
-	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5, lwd=2)
-	for (i in seq_len(length(indices)+1))
-	{
-		ind = ifelse(i<=length(indices), indices[i] - first_day + 1, index)
-		plot(c(data$getCenteredSerie(ind),data$getCenteredSerie(ind+1)),
-			ylim=yrange, type="l", col=colors[i],
-			xlab=ifelse(i==1,"Temps (en heures)",""), ylab=ifelse(i==1,"PM10 centré",""))
-		if (i <= length(indices))
-			par(new=TRUE)
-	}
+	prediction = pred$getForecast(index)
+	measure = data$getSerie( pred$getIndexInData(index) )[1:length(pred$getForecast(1))]
+
+	# Remove the common part, where prediction == measure
+	dot_mark <- ifelse(prediction[1]==measure[1],
+		which.max(seq_along(prediction)[prediction==measure]), 0)
+	prediction = prediction[(dot_mark+1):length(prediction)]
+	measure = measure[(dot_mark+1):length(measure)]
+
+	yrange = range(measure, prediction)
+	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5, lwd=3)
+	plot(measure, type="l", ylim=yrange, xlab="Time (hours)", ylab="PM10")
+	par(new=TRUE)
+	plot(prediction, type="l", col="#0000FF", ylim=yrange, xlab="", ylab="")
 }
 
-#' @title Plot similarities
+#' Plot similarities
 #'
-#' @description Plot histogram of similarities (weights)
+#' Plot histogram of similarities (weights), for 'Neighbors' method.
 #'
-#' @param pred Object as returned by \code{getForecast}
-#' @param index Index in forecasts (not in data)
+#' @inheritParams computeError
+#' @param index Index in forecasts (integer or date)
 #'
 #' @export
 plotSimils <- function(pred, index)
@@ -102,94 +104,160 @@ plotSimils <- function(pred, index)
 	weights = pred$getParams(index)$weights
 	if (is.null(weights))
 		stop("plotSimils only works on 'Neighbors' forecasts")
+	par(mfrow=c(1,2), mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
+	small_weights = weights[ weights < 1/length(weights) ]
+	large_weights = weights[ weights >= 1/length(weights) ]
+	hist(small_weights, nclass=25, main="", xlab="Weight < 1/N", ylab="Count")
+	hist(large_weights, nclass=25, main="", xlab="Weight >= 1/N", ylab="Count")
+}
+
+#' Functional boxplot
+#'
+#' Draw the functional boxplot on the left, and bivariate plot on the right.
+#'
+#' @inheritParams computeError
+#' @inheritParams plotCurves
+#'
+#' @export
+plotFbox <- function(data, indices=seq_len(data$getSize()))
+{
+	if (!requireNamespace("rainbow", quietly=TRUE))
+		stop("Functional boxplot requires the rainbow package")
+
+	series_matrix = data$getSeries(indices)
+	# Remove series with NAs
+	no_NAs_indices = sapply( 1:ncol(series_matrix),
+		function(i) all(!is.na(series_matrix[,i])) )
+	series_matrix = series_matrix[,no_NAs_indices]
+
+	series_fds = rainbow::fds(seq_len(nrow(series_matrix)), series_matrix)
 	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
-	hist(pred$getParams(index)$weights, nclass=20, xlab="Poids", ylab="Effectif")
+	rainbow::fboxplot(series_fds, "functional", "hdr", xlab="Time (hours)", ylab="PM10",
+		plotlegend=FALSE, lwd=2)
+	rainbow::fboxplot(series_fds, "bivariate", "hdr", plotlegend=FALSE)
 }
 
-#' @title Plot error
+#' Compute filaments
 #'
-#' @description Draw error graphs, potentially from several runs of \code{getForecast}
+#' Obtain similar days in the past, and (optionally) plot them -- as black as distances
+#' are small.
 #'
-#' @param err Error as returned by \code{getError}
-#' @param cols Colors for each error (default: 1,2,3,...)
+#' @inheritParams computeError
+#' @param index Index in forecast (integer or date)
+#' @param limit Number of neighbors to consider
+#' @param plot Should the result be plotted?
 #'
-#' @seealso \code{\link{plotPredReal}}, \code{\link{plotFilaments}}, \code{\link{plotSimils}}
-#'   \code{\link{plotFbox}}
+#' @return A list with
+#' \itemize{
+#'   \item index : index of the current serie ('today')
+#'   \item neighb_indices : indices of its neighbors
+#'   \item colors : colors of neighbors curves (shades of gray)
+#' }
 #'
 #' @export
-plotError <- function(err, cols=seq_along(err))
+computeFilaments <- function(data, pred, index, limit=60, plot=TRUE)
 {
-	if (!is.null(err$abs))
-		err = list(err)
-	par(mfrow=c(2,2), mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5, lwd=2)
-	L = length(err)
-	yrange = range( sapply(1:L, function(index) ( err[[index]]$abs$day ) ), na.rm=TRUE )
-	for (i in seq_len(L))
-	{
-		plot(err[[i]]$abs$day, type="l", xlab=ifelse(i==1,"Temps (heures)",""),
-			ylab=ifelse(i==1,"Moyenne |y - y_hat|",""), ylim=yrange, col=cols[i])
-		if (i < L)
-			par(new=TRUE)
-	}
-	yrange = range( sapply(1:L, function(index) ( err[[index]]$abs$indices ) ), na.rm=TRUE )
-	for (i in seq_len(L))
-	{
-		plot(err[[i]]$abs$indices, type="l", xlab=ifelse(i==1,"Temps (jours)",""),
-			ylab=ifelse(i==1,"Moyenne |y - y_hat|",""), ylim=yrange, col=cols[i])
-		if (i < L)
-			par(new=TRUE)
-	}
-	yrange = range( sapply(1:L, function(index) ( err[[index]]$MAPE$day ) ), na.rm=TRUE )
-	for (i in seq_len(L))
+	weights <- pred$getParams(index)$weights
+	if (is.null(weights) || is.na(pred$getParams(index)$weights[1]))
+		stop("computeFilaments requires a serie without NAs")
+
+	nn <- min(limit, length(weights))
+	sorted_dists = sort(-log(weights), index.return=TRUE)
+	# Compute colors for each neighbor (from darkest to lightest), if weights differ
+	if ( any( weights != weights[1] ) )
 	{
-		plot(err[[i]]$MAPE$day, type="l", xlab=ifelse(i==1,"Temps (heures)",""),
-			ylab=ifelse(i==1,"MAPE moyen",""), ylim=yrange, col=cols[i])
-		if (i < L)
-			par(new=TRUE)
+		min_dist = min(sorted_dists$x[1:nn])
+		max_dist = max(sorted_dists$x[1:nn])
+		color_values = floor(19.5*(sorted_dists$x[1:nn]-min_dist)/(max_dist-min_dist)) + 1
+		colors = gray.colors(20,0.1,0.9)[color_values] #TODO: 20 == magic number
 	}
-	yrange = range( sapply(1:L, function(index) ( err[[index]]$MAPE$indices ) ), na.rm=TRUE )
-	for (i in seq_len(L))
+	else
+		colors <- rep(colors()[17], length(weights))
+
+	if (plot)
 	{
-		plot(err[[i]]$MAPE$indices, type="l", xlab=ifelse(i==1,"Temps (jours)",""),
-			ylab=ifelse(i==1,"MAPE moyen",""), ylim=yrange, col=cols[i])
-		if (i < L)
+		# Complete series with (past and present) tomorrows
+		ref_serie = c( data$getCenteredSerie( pred$getIndexInData(index)-1 ),
+			data$getCenteredSerie( pred$getIndexInData(index) ) )
+		centered_series = rbind(
+			data$getCenteredSeries( pred$getParams(index)$indices-1 ),
+			data$getCenteredSeries( pred$getParams(index)$indices ) )
+		yrange = range( ref_serie,
+			quantile(centered_series, probs=c(0.025,0.975), na.rm=TRUE) )
+		par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5, lwd=2)
+		for (i in nn:1)
+		{
+			plot(centered_series[,sorted_dists$ix[i]], ylim=yrange, type="l", col=colors[i],
+				xlab=ifelse(i==1,"Time (hours)",""), ylab=ifelse(i==1,"Centered PM10",""))
 			par(new=TRUE)
+		}
+		# Also plot ref curve, in red
+		plot(ref_serie, ylim=yrange, type="l", col="#FF0000", xlab="", ylab="")
+		dot_mark <- 0.5 + which.max( pred$getForecast(1) ==
+			data$getSerie( pred$getIndexInData(1) )[1:length(pred$getForecast(1))] )
+		abline(v=24+dot_mark, lty=2, col=colors()[56], lwd=1)
 	}
+
+	list(
+		"index"=pred$getIndexInData(index),
+		"neighb_indices"=pred$getParams(index)$indices[sorted_dists$ix[1:nn]],
+		"colors"=colors)
 }
 
-#' @title Functional boxplot
+#' Functional boxplot on filaments
 #'
-#' @description Draw the functional boxplot on the left, and bivariate plot on the right
+#' Draw the functional boxplot on filaments obtained by \code{computeFilaments()}.
 #'
-#' @param data Object return by \code{getData}
-#' @param fiter Optional filter: return TRUE on indices to process
+#' @inheritParams computeError
+#' @param fil Output of \code{computeFilaments}
+#' @param predict_from First predicted time step
 #'
 #' @export
-plotFbox <- function(data, filter=function(index) TRUE)
+plotFilamentsBox = function(data, fil, predict_from)
 {
 	if (!requireNamespace("rainbow", quietly=TRUE))
 		stop("Functional boxplot requires the rainbow package")
 
-	start_index = 1
-	end_index = data$getSize()
-	if (length(data$getCenteredSerie(1)) < length(data$getCenteredSerie(2)))
-	{
-		# Shifted start (7am, or 1pm, or...)
-		start_index = 2
-		end_index = data$getSize() - 1
-	}
-
-	series_matrix = sapply(start_index:end_index, function(index) {
-		as.matrix(data$getSerie(index))
-	})
-	# Remove NAs. + filter TODO: merge with previous step: only one pass required...
-	nas_indices = seq_len(ncol(series_matrix))[ sapply( 1:ncol(series_matrix),
-		function(index) ( !filter(index) || any(is.na(series_matrix[,index])) ) ) ]
-	series_matrix = series_matrix[,-nas_indices]
-
+	series_matrix = rbind(
+		data$getSeries(fil$neighb_indices-1), data$getSeries(fil$neighb_indices) )
 	series_fds = rainbow::fds(seq_len(nrow(series_matrix)), series_matrix)
-	par(mfrow=c(1,2), mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
-	rainbow::fboxplot(series_fds, "functional", "hdr", xlab="Temps (heures)", ylab="PM10",
+
+	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
+	rainbow::fboxplot(series_fds, "functional", "hdr", xlab="Time (hours)", ylab="PM10",
 		plotlegend=FALSE, lwd=2)
-	rainbow::fboxplot(series_fds, "bivariate", "hdr", plotlegend=FALSE)
+
+	# "Magic": http://stackoverflow.com/questions/13842560/get-xlim-from-a-plot-in-r
+	usr <- par("usr")
+	yr <- (usr[4] - usr[3]) / 27
+	par(new=TRUE)
+	plot(c(data$getSerie(fil$index-1),data$getSerie(fil$index)), type="l", lwd=2, lty=2,
+		ylim=c(usr[3] + yr, usr[4] - yr), xlab="", ylab="")
+	abline(v=24+predict_from-0.5, lty=2, col=colors()[56])
+}
+
+#' Plot relative conditional variability / absolute variability
+#'
+#' Draw the relative conditional variability / absolute variability based on filaments
+#' obtained by \code{computeFilaments()}.
+#'
+#' @inheritParams computeError
+#' @inheritParams plotFilamentsBox
+#'
+#' @export
+plotRelVar = function(data, fil, predict_from)
+{
+	ref_var = c( apply(data$getSeries(fil$neighb_indices-1),1,sd),
+		apply(data$getSeries(fil$neighb_indices),1,sd) )
+	tdays = .getNoNA2(data, 2, fil$index)
+	global_var = c(
+		apply(data$getSeries(tdays-1),1,sd),
+		apply(data$getSeries(tdays),1,sd) )
+
+	yrange = range(ref_var, global_var)
+	par(mar=c(4.7,5,1,1), cex.axis=1.5, cex.lab=1.5)
+	plot(ref_var, type="l", col=1, lwd=3, ylim=yrange,
+		xlab="Time (hours)", ylab="Standard deviation")
+	par(new=TRUE)
+	plot(global_var, type="l", col=2, lwd=3, ylim=yrange, xlab="", ylab="")
+	abline(v=24+predict_from-0.5, lty=2, col=colors()[56])
 }