Modelling the number of reverse dependencies
2023-08-17
From the dependency network of all CRAN packages, we have seen that reverse dependencies of all CRAN packages seem to follow the power law. This echoes the phenomenon in other software dependencies observed by, for example, LaBelle and Wallingford, 2004, Baxter et al., 2006, Jenkins and Kirk, 2007, Wu et al. 2007, Louridas et al., 2008, Zheng et al., 2008, Kohring, 2009, Li et al., 2013, Bavota et al. 2015, and Cox et al., 2015. In this vignette, we will fit the discrete power law to model this number of reverse dependencies, using functions with the suffix upp, namely dupp(), Supp() and mcmc_upp(). While we shall focus on “Imports”, which is one of the serveral kinds of dependencies in R, the same analysis can be carried out for all other types.
Obtaining the reverse dependencies
In the vignette on all CRAN pacakges, we looked at the dependency type Depends. Here, we look at the dependency type Imports, using the same workflow. As we are using the forward dependencies to construct the network, the number of reverse dependencies is equivalent to the in-degree of a package, which can be obtained via the function igraph::degree(). We then construct a data frame to hold this in-degree information.
For the purpose of verification, we use the reverse dependencies to construct the network, then look at the out-degrees of the packages this time.
Theoretically, the two data frames are identical. Any possible (but small) difference is due to the CRAN pages being updated while scraping.
Exploratory analysis and selecting the threshold
We construct a data frame for the empirical frequencies and survival function at the whole range of data.
Before fitting the discrete power law, to be determined first is a threshold above which it is appropriate. We will visualise the degree distribution to determine such threshold.
The power law seems appropriate for the whole range of data, and so, for illustration purposes, the threshold will be set at 1 inclusive. 0’s will be excluded anyway because, as we will see, the probability mass function (PMF) is not well-defined at 0.
While it is straightforward to determine the threshold here, such linearity over the whole range might not be seen for other data. The package poweRlaw provides functions for and references to more systematic/objective procedures of selecting the threshold.
Fitting the discrete power law
We use the function mcmc_pol() to fit the discrete power law, of which the PMF is proportional to \(x^{-\alpha}\), where \(\alpha\) is the lone scalar parameter.
The Bayesian approach is used here for inference, meaning that a prior has to be set for the parameters. We assume a normal distribution \(N(m_{\alpha}=0, s_{\alpha}=10)\) for \(\alpha\). Markov chain Monte Carlo (MCMC) is used as the inference algorithm.
x <- df1.imports$degree
counts <- df1.imports$frequency
alpha0 <- 1.5 # initial value
theta0 <- 1.0 # initial value, but fixed to 1.0 for discrete power law
m_alpha <- 0.0 # prior mean
s_alpha <- 10.0 # prior standard deviation
set.seed(3075L)
mcmc0.imports <-
mcmc_pol(
x = x,
count = counts,
alpha = alpha0,
theta = theta0,
a_alpha = m_alpha,
b_alpha = s_alpha,
a_theta = 1.0,
b_theta = 1.0,
a_pseudo = 10.0,
b_pseudo = 1.0,
pr_power = 1.0,
iter = 2500L,
thin = 1L,
burn = 500L,
freq = 500L,
invt = 1.0,
mc3_or_marg = TRUE,
x_max = 100000
) # takes seconds
Now we have the samples representing the posterior distribution of \(\alpha\):
This means the number of reverse “Imports” follows approximately a power law with exponent 1.65. We can also obtain the posterior density via numerical integration using marg_pol(). The returned list contains log.marginal for the marginal log-likelihood, and posterior for the posterior density in a data frame. We verify its alignment with the MCMC results:
We can also calculate the fitted frequencies and survival function, using dpol() and Spol() respectively. Their summaries are however readily available in the MCMC output. We overlay the fitted line (blue, dashed) and credible intervals (red, dotted) on the plot above to check goodness-of-fit:
n0 <- sum(df1.imports$frequency) # TOTAL number of data points in x
## or n0 <- length(x)
df0.fitted <- mcmc0.imports$fitted
gg1 <- df1.imports |>
plot_log_log() +
ggplot2::geom_point(aes(degree, frequency), size = 0.75) +
ggplot2::geom_line(aes(x, f_med * n0), data = df0.fitted, col = 4, lty = 2) +
ggplot2::geom_line(aes(x, f_025 * n0), data = df0.fitted, col = 2, lty = 3) +
ggplot2::geom_line(aes(x, f_975 * n0), data = df0.fitted, col = 2, lty = 3) +
ggplot2::coord_cartesian(ylim = c(1L, 2e+3L))
gg1
Fitting extreme value mixture distribution
The corresponding survival function according to the power law fit can also be obtained:
gg2 <- df1.imports |>
plot_log_log() +
ggplot2::geom_point(aes(degree, survival), size = 0.75) +
ggplot2::geom_line(aes(x, S_med), data = df0.fitted, col = 4, lty = 2) +
ggplot2::geom_line(aes(x, S_025), data = df0.fitted, col = 2, lty = 3) +
ggplot2::geom_line(aes(x, S_975), data = df0.fitted, col = 2, lty = 3)
gg2
#> Warning in ggplot2::scale_y_log10(): log-10 transformation introduced infinite
#> values.
This shows the discrete power law doesn’t fit as good as it seems according to the frequency plot. To improve the fit, we use the discrete extreme value mixture distributions introduced in Lee, Eastoe and Farrell, 2024. Provided in this package for such purpose are dmix2(), Smix2() and mcmc_mix2(). As it takes hours to run the inference algorithm using mcmc_mix(), we hardcode the parameter values based on their component-wise posterior median, and overlay the data by the fitted line.
gg3 <- df1.imports |>
filter(degree > 1L) |>
dplyr::mutate(
survival = 1.0 - cumsum(frequency) / sum(frequency),
survival.mix = Smix2(degree, 1606, 1.73, 1.00, 0.237, 4.03, 0.003)
) |>
ggplot2::ggplot() +
ggplot2::geom_point(aes(degree, survival), size = 0.75) +
ggplot2::geom_line(aes(degree, survival.mix), col = 4, lty = 2, lwd = 1.2) +
ggplot2::scale_x_log10() +
ggplot2::scale_y_log10() +
ggplot2::theme_bw(12)
gg3
#> Warning in ggplot2::scale_y_log10(): log-10 transformation introduced infinite
#> values.
This illustrates the mixture distribution’s better fit.