Welcome to the mvrsquared
package! This package does one
thing: calculate the coefficient
of determination or R-squared. However, this implementation is
different from what you may be familiar with. In addition to the
standard R-squared used frequently in linear regression,
mvrsquared
calculates R-squared for multivariate outcomes.
(This is why there is an ‘mv’ in mvrsquared
).
mvrsquared
implements R-squared based on a derivation in
this paper. It’s the same
definition of R-squared you’re probably familiar with (\(1 - \frac{SSE}{SST}\)) but generalized to
n-dimensions.
In the standard case, your outcome \(y\) and prediction \(\hat{y}\) are vectors. In other words, each observation is a single number. This is fine if you are predicting a single variable. But what if you are predicting multiple variables at once? In that case, \(y\) and \(\hat{y}\) are matrices. This situation occurs frequently in topic modeling or simultaneous equation modeling.
Below are some examples of using mvrsquared
for
calculating R-squared in the traditional context and for multivariate
outcomes.
Below is a simple example comparing the R-squared calculated from a
linear model fit by lm
in the stats
package
and the R-squared calculated in the mvrsquared
package.
(Spoiler alert: they’re identical.) Why do this? If
mvrsquared
can’t get the same calculation, you probably
shouldn’t trust it. (Spoiler alert again: it can and you should.)
library(mvrsquared)
data(mtcars)
# fit a linear model
f <- lm(mpg ~ cyl + disp + hp + wt, data = mtcars)
# extract r-squared
f_summary <- summary(f)
r2_lm <- f_summary$r.squared
r2_lm
#> [1] 0.8486348
# calculate univariate r-squared using mvrsquared
r2_mv <- calc_rsquared(y = mtcars$mpg, yhat = f$fitted.values)
r2_mv
#> [1] 0.8486348
# just to be 100% sure...
r2_lm == r2_mv
#> [1] TRUE
You can also calculate R-squared by passing your data and your linear weights. (That’s probably less useful for univariate outcome models. But, as you’ll see below, it’s very important for latent variable models like topic models.)
x <- cbind(1, f$model[, -1]) # note, you have to add 1's for the intercept and
# I'm removing the first column of f$model as it
# is the outcome we are predicting
x <- as.matrix(x) # x needs to be a matrix, not a data.frame or tibble for now
w <- matrix(f$coefficients, ncol = 1) # w also has to be a matrix
# this calculates yhat as the dot product x %*% w
r2_mv2 <- calc_rsquared(y = mtcars$mpg,
yhat = list(x = x,
w = w))
r2_mv2
#> [1] 0.8486348
Calculating R-squared this way does lead to a tiny difference in calculation due to numeric precision.
However, the difference is tiny. Below demonstrates that they are the same up to 14 decimal places in this example.
Below, just to show you I’m not blowing smoke, is a silly example that demonstrates that you don’t get any crazy results just because your outcome is a matrix. If we make the columns of \(y\) the same values (and ditto for \(\hat{y}\)), you actually get the same R-squared as above.
calc_rsquared(y = cbind(mtcars$mpg, mtcars$mpg),
yhat = cbind(f$fitted.values, f$fitted.values))
#> [1] 0.8486348
Same as above, no?
Here’s a more realistic example. Say you’ve got a neural net that is predicting two variables at once.
library(nnet)
# let's generate some synthetic data
set.seed(666)
# Some continuous variables
x1 <- rnorm(n = 10000, mean = 1, sd = 2)
x2 <- rnorm(n = 10000, mean = 1.5, sd = 2.5)
x3 <- rnorm(n = 10000, mean = .6, sd = 1.5)
# linear combinations used to generate outcomes with logit functions
z1 <- 1 - 1.2 * x1 + 4.5 * x2 - 2.8 * x3
z2 <- -2 + 2.8 * x1 - 1.2 * x2 + 4.5 * x3
y1 <- rbinom(10000, 1, prob = 1 / (1 + exp(-z1)))
y2 <- rbinom(10000, 1, prob = 1 / (1 + exp(-z2)))
# fit a multinomial model using a 1-layer neural net with 10 nodes
f_mv <- nnet(cbind(y1, y2) ~ x1 + x2 + x3,
size = 10)
#> # weights: 62
#> initial value 6406.585738
#> iter 10 value 3246.812009
#> iter 20 value 1446.678774
#> iter 30 value 1060.315445
#> iter 40 value 1004.698287
#> iter 50 value 961.043723
#> iter 60 value 932.089225
#> iter 70 value 907.861456
#> iter 80 value 901.751981
#> iter 90 value 899.029455
#> iter 100 value 896.508832
#> final value 896.508832
#> stopped after 100 iterations
yhat <- predict(f_mv, data.frame(x1 = x1, x2 = x2, x3 = x3), type = "raw")
# and now calculate r-squared
calc_rsquared(y = cbind(y1, y2), yhat = yhat)
#> [1] 0.8094835
The paper (as of this writing still in progress) that derives the multivariate R-squared is focused on probabilistic topic modeling. Here’s an example using Latent Dirichlet Allocation.
Below are preliminaries of loading data and getting the model.
library(tidytext)
library(textmineR)
#> Loading required package: Matrix
#>
#> Attaching package: 'textmineR'
#> The following object is masked from 'package:Matrix':
#>
#> update
#> The following object is masked from 'package:stats':
#>
#> update
library(dplyr)
#>
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#>
#> filter, lag
#> The following objects are masked from 'package:base':
#>
#> intersect, setdiff, setequal, union
library(stringr)
# load documents in a data frame
docs <- nih_sample
# tokenize using tidytext's unnest_tokens
tidy_docs <- docs %>%
select(APPLICATION_ID, ABSTRACT_TEXT) %>%
unnest_tokens(output = word,
input = ABSTRACT_TEXT,
stopwords = stop_words$word,
token = "ngrams",
n_min = 1, n = 2) %>%
count(APPLICATION_ID, word) %>%
filter(n>1) #Filtering for words/bigrams per document, rather than per corpus
tidy_docs <- tidy_docs %>% # filter words that are just numbers
filter(! str_detect(tidy_docs$word, "^[0-9]+$"))
# turn a tidy tbl into a sparse dgCMatrix for use in textmineR
dtm <- tidy_docs %>%
cast_sparse(APPLICATION_ID, word, n)
# create a topic model
lda <- FitLdaModel(dtm = dtm,
k = 20,
iterations = 200,
burnin = 175)
The paper notes that for a probabilistic topic model, \(\hat{y} = \vec{n} \odot \Theta \cdot \Phi\), where \(\vec{n}\) is a vector of document lengths, \(\odot\) is elementwise multiplication, \(\Theta\) is a matrix of topic distributions over documents, and \(\Phi\) is a matrix of token distributions over topics.
Doing this multiplication explicitly in R could lead you to run out
of RAM for a decently-sized corpus. (This toy example won’t.) Luckily,
we can use the ability to pass a list x
and w
to calc_rsquared
to do that multiplication more
efficiently. In that case, we set x
\(= \vec{n}\odot\Theta\) and w
\(= \Phi\).
(Note that you can pass a sparse dgCMatrix
from the
Matrix
or something coercible to one directly to
calc_rsquared
.)
Latent semantic analysis (LSA) uses a single value decomposition of a document term matrix as a non-probabilistic topic model.
Note the example below doesn’t use TF-IDF even
though most folks recommend you do. For a more thorough example of LSA
(and LDA) see textmineR
’s vignette
on topic modeling.
You say you have BIG DATA?!?! Below is an example of how to chop up a
calculation and have r-squared built in parallel in a distributed
setting. mvrsquared
has a parallel back end using RcppThread
and controlled by the threads
argument. However, if that
isn’t enough, you can bring your own distributed parallel framework. The
example below uses parallel::mclapply
which will not
parallelize on Windows, but R has other options.
mvrsquared
has the option to return only the sums of
squares (specifically SSE and SST). You can calculate these in
parallel/distributed and combine them by adding post-hoc. (That’s a
little algorithm called “map-reduce” for those in the know.) Once
combined, you can calculate R-squared yourself with the canonical
formula \(1 - \frac{SSE}{SST}\).
Note: You must calculate the mean of your outcome,
\(\bar{y}\), first and hand it off to
calc_rsquared
. Otherwise, your calculations for SST on a
given core will be based only on the observations sent to that core, not
the whole data set. (If you do this, calc_rsquared
will
give you a WARNING.)
Below is a toy example using parallel::mclapply
to run
parallel computations calculating R-squared for our LDA model, above.
(Yes, this is overkill for our toy problem and it will actually be
slower due to overhead. But imagine the calculations for a corpus of ~1
million documents with hundreds-of-thousands or millions of unique
tokens.)
Because the below example uses parallel execution, which can be tricky when working with CRAN, I’m not actually evaluating it in this vignette. However, you can run the code yourself.
library(parallel)
batch_size <- 10
batches <- mclapply(X = seq(1, nrow(dtm), by = batch_size),
FUN = function(b){
# rows to select on
rows <- b:min(b + batch_size - 1, nrow(dtm))
# rows of the dtm
y_batch <- dtm[rows, ]
# rows of theta multiplied by document length
x_batch <- rowSums(y_batch) * lda$theta[rows, ]
list(y = y_batch,
x = x_batch)
}, mc.cores = 2)
# calculate ybar for the data
# in this case, lazily doing colMeans, but you could divide this problem up too
ybar <- colMeans(dtm)
# MAP: calculate sums of squares
ss <- mclapply(X = batches,
FUN = function(batch){
calc_rsquared(y = batch$y,
yhat = list(x = batch$x, w = lda$phi),
ybar = ybar,
return_ss_only = TRUE)
}, mc.cores = 2)
# REDUCE: get SST and SSE by summation
ss <- do.call(rbind, ss) %>% colSums()
r2_mapreduce <- 1 - ss["sse"] / ss["sst"]
# should be the same as above
r2_mapreduce