Heatmaps with eye-tracking data (weighted 2D-density)

user_1969717 0
Viewed 121

I am trying to create a fixation maps where the weight of each fixation on the 2d density map is determined by its duration. As I understood, the stat_density2d() function accepts the weight argument but does not process it (ggplot2 2d Density Weights)

Is there a way to go around this ? Also, how could I smooth the granularity of the heatmaps? I must be missing something quite obvious here

#sample data
set.seed(42)  ## for sake of reproducibility
df <- data.frame(x=sample(0:1920, 1000, replace=TRUE), 
                 y=sample(0:1080, 1000, replace=TRUE), 
                 dur=sample(50:1000, 1000, replace=TRUE))

#what I have so far
library(ggplot2)
ggplot(df, aes(x=x, y =y)) +
  stat_density2d(geom='raster', 
                 aes(fill=..count.., alpha=..count..), contour=FALSE) + 
  geom_point(aes(size=dur), alpha=0.2, color="red") +
  scale_fill_gradient(low="green", high="red") +
  scale_alpha_continuous(range=c(0, 1) , guide="none") +
  theme_void()

enter image description here

heatmap ggplot2 r
2 Answers

Not a ggplot2 user, but basically you want to estimate a weighted 2d-density and make an image out of it. Your linked answer indicates that ggplot2::geom_density2d internally uses MASS::kde2d, but which only computes unweighted 2d-densities.

Inflate observations

Similar to @AllanCameron's suggestion (but without the need to use tidyr) we could inflate the data frame simply by copying each row by the number of milliseconds duration,

dfa <- df[rep(seq_len(nrow(df)), times=df$dur), -3]

and calculate the kde2d by hand.

n <- 1e3

system.time(
  dens1 <- MASS::kde2d(dfa$x, dfa$y, n=n)  ## this runs a while!
)
#     user   system  elapsed 
# 2253.285 2325.819  661.632

The n= argument denotes the number of grid points in each direction, the greater we choose it, the smoother the granularity will look in the heatmap image.

system.time(
  dens1 <- MASS::kde2d(dfa$x, dfa$y, n=n)  ## this runs a while
)
#     user   system  elapsed 
# 2253.285 2325.819  661.632 

image(dens1, col=heat.colors(n, rev=TRUE))

enter image description here

This almost runs forever, though with n=1000...

Weighted 2d density estimation

In a comment on the answer mentioned above, @IRTFM links an ancient r-help post that provides a kde2d.weighted function which is lightning fast and that we could try (see code at the bottom).

dens2 <- kde2d.weighted(x=df$x, y=df$y, w=proportions(df$dur), n=n) 
image(dens2, col=heat.colors(n, rev=TRUE))

enter image description here

However, the two versions look quite different, and I can't tell which is right, since I am not really an expert with this method. But at least there is a noticeable difference to the unweighted image:

Unweighted image

dens0 <- MASS::kde2d(df$x, df$y, n=n)
image(dens0, col=heat.colors(n, rev=TRUE))

enter image description here

Points

Still adding the points might be pointless, but you could run this line after image:

points(y ~ x, df, cex=proportions(dur)*2e3, col='green')

Taken from r-help (Ort 2006):

kde2d.weighted <- function(x, y, w, h, n=n, lims=c(range(x), range(y))) {
  nx <- length(x)
  if (length(y) != nx) 
    stop("data vectors must be the same length")
  gx <- seq(lims[1], lims[2], length=n)  ## gridpoints x
  gy <- seq(lims[3], lims[4], length=n)  ## gridpoints y
  if (missing(h)) 
    h <- c(MASS::bandwidth.nrd(x), MASS::bandwidth.nrd(y))
  if (missing(w)) 
    w <- numeric(nx) + 1
  h <- h/4
  ax <- outer(gx, x, "-")/h[1]  ## distance of each point to each grid point in x-direction
  ay <- outer(gy, y, "-")/h[2]  ## distance of each point to each grid point in y-direction
  z <- (matrix(rep(w,n), nrow=n, ncol=nx, byrow=TRUE)*
          matrix(dnorm(ax), n, nx)) %*% 
    t(matrix(dnorm(ay), n, nx))/(sum(w)*h[1]*h[2])  ## z is the density
  return(list(x=gx, y=gy, z=z))
}
user_65740380

The simplest way round this is to use tidyr::uncount to replicate the rows of your data frame, using dur as weights:

library(ggplot2)

ggplot(tidyr::uncount(df, dur), aes(x=x, y =y)) +
  stat_density2d(geom='raster', 
                 aes(fill=..count.., alpha=..count..), contour=FALSE) + 
  geom_point(data = df, aes(size=dur), alpha=0.2, color="red") +
  scale_fill_gradient(low="green", high="red") +
  scale_alpha_continuous(range=c(0, 1) , guide="none") +
  theme_void()

enter image description here

The effect is probably easier to see with the points removed:

ggplot(tidyr::uncount(df, dur), aes(x=x, y =y)) +
  stat_density2d(geom='raster', 
                 aes(fill=..count.., alpha=..count..), contour=FALSE) + 
  scale_fill_gradient(low="green", high="red") +
  scale_alpha_continuous(range=c(0, 1) , guide="none") +
  theme_void()

enter image description here

user_125003150
Related