The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.

Introduction

The notion of spreading activation is a prevalent metaphor in the cognitive sciences; however, the tools to implement spreading activation in a computational simulation are not as readily available. This vignette introduces the spreadr R package (pronunced ‘SPREAD-er’), which can implement spreading activation within a specified network structure. The algorithmic method implemented in the spreadr function follows the approach described in Vitevitch, Ercal, and Adagarla (2011), who viewed activation as a fixed cognitive resource that could “spread” among connected nodes in a network.

Installation

You can choose to install the stable version from CRAN, or the development build from GitHub.

Stable (CRAN)

install.packages("spreadr")

If you encounter any bugs or issues, please try the development build first. The bug or issue may have already been fixed on GitHub, but not yet propagated onto CRAN.

Development (GitHub)

install.packages("remotes")
remotes::install_github("csqsiew/spreadr")

If you encounter any bugs or issues, please try the development build first. The bug or issue may have already been fixed on GitHub, but not yet propagated onto CRAN.

Example with a phonological network

In this example, we will simulate spreading activation in a small sample portion of a phonological network (Chan and Vitevitch 2009) which is automatically loaded with spreadr. This phonological network is unweighted and undirected, but spreadr supports weighted and directed graphs as well. This makes it is possible to simulate spreading activation in a weighted network where more activation is passed between nodes that have “stronger” edges, or in a directed (asymmetric) network where activation can pass from node i to node j but not necessarily from node j to node i.

You may substitute any network in place of this example one.

Describe the network

The network for spreading activation must be either an igraph object or an adjacency matrix.

Using an igraph object

pnet, an igraph object with named vertices representing our sample phonological network is automatically loaded with the spreadr library.

library(spreadr)
## Warning: package 'Rcpp' was built under R version 4.0.2
library(igraph)
## Warning: package 'igraph' was built under R version 4.0.2
data("pnet")  # load and inspect the igraph object
plot(pnet)

Don’t worry if your plot does not look exactly as above. The layout of the nodes is determined stochastically.

Using an adjacency matrix

pnetm, an named adjacency matrix representing our sample phonological network is automatically loaded with the spreadr library.

library(spreadr)
library(igraph)
set.seed(1)

data("pnetm")
pnetm[1:5, 1:5]  # inspect the first few entries
##       spike speak spoke speck spook
## spike     0     1     1     1     1
## speak     1     0     1     1     1
## spoke     1     1     0     1     1
## speck     1     1     1     0     1
## spook     1     1     1     1     0
plot(graph_from_adjacency_matrix(
  pnetm, mode="undirected")) # visualise the graph

For those following along: don’t worry if your plot does not look exactly as above. The layout of the nodes is determined stochastically.

For simplicity, the rest of this example will use pnet, an igraph representation of pnetm. This difference is trivial — the spreadr function accepts both igraph and adjacency matrix.

Specify the adding of activation

A simulation of spreading activation is uninteresting without some activation to spread. In the simplest case, you may specify the initial activation state of each node, from which the simulation will proceed. Alternatively, you can also specify the addition of activation at any node, at any time point within the simulation.

Initial activation only

Initial activation is specified by a data.frame (or data.frame-like object, such as a tibble) with columns node and activation. Each row represents the addition of activation amount of activation to the node with name node, at the initial pre-spreading activation step of the simulation.

For example, if we wanted the nodes "beach" and "speck" to have 20 and 10 activation initially, respectively:

start_run <- data.frame(
  node=      c("beach", "speck"),
  activation=c(     20,      10))

Activation at specified time points

The adding of activation at specified time points is specified by a data.frame (or data.frame-like object, such as a tibble) with columns node, activation, and time. Each row represents the addition of activation amount of activation to the node with name node at time point time. The initial state before any spreading activation occurs is time = 0.

Therefore, if we wanted the node "beach" to have 20 activation initially, then "speck" to have 10 activation at time = 5:

start_run <- data.frame(
  node =       c("beach", "speck"),
  activation = c(     20,      10),
  time =       c(      0,       5))

For more details about the meaning of time in the spreading activation simulation, please read the spreadr function documentation. To keep things simple, the rest of this example will use the version of start_run without the time column (see the leftmost tab).

Run the simulation

We are now ready to run the simulation through the spreadr function. This function takes a number of parameters, some of which are listed here. Required parameters are written in bold, optional parameters are written with their default values (in brackets):

  1. network: Adjacency matrix or igraph object representing the network in which to simulate spreading activation
  2. start_run: Non-empty data.frame describing the addition of activation into the network.
  3. decay (0): Proportion of activation lost at each time step (range from 0 to 1)
  4. retention (0.5): Proportion of activation retained in the originator node (range from 0 to 1)
  5. suppress (0): Nodes with activation values lower than this value will have their activations forced to 0 (typically this will be a very small value e.g. < 0.001)
  6. time (10): Number of iterations in the simulation
  7. include_t0 (FALSE): If the initial state activations should be included in the result data.frame

For more details on all the possible function parameters, please read the spreadr function documentation. For more details on the spreading activation algorithm, see Siew (2019) and Vitevitch, Ercal, and Adagarla (2011).

result <- spreadr(pnet, start_run, include_t0=TRUE)

Examine the results

The result is presented as a data.frame with columns node, activation, and time. Each row contains the activation value of a node at its specified time step.

head(result)  # inspect the first few rows
##    node activation time
## 1 spike          0    0
## 2 speak          0    0
## 3 spoke          0    0
## 4 speck         10    0
## 5 spook          0    0
## 6  sped          0    0
tail(result)  # inspect the last few rows
##       node activation time
## 369  patch  0.5139104   10
## 370  pooch  0.5139104   10
## 371  poach  0.5139104   10
## 372 preach  0.5493873   10
## 373  reach  2.0688303   10
## 374  perch  0.5139104   10

As a data.frame, the result can be easily saved as a CSV file for sharing.

write.csv(result, file="result.csv")  # save result as CSV file

You can also easily visualise the result using your favourite graphical libraries. Here, we use ggplot2:

library(ggplot2)
## Warning: package 'ggplot2' was built under R version 4.0.2
ggplot(result, aes(x=time, y=activation, color=node)) +
  geom_point() +
  geom_line()

Additional examples

Here, we address some common “how do I do X?” questions by example.

Using a weighted network

Simply pass a weighted network (igraph object or adjacency matrix) to the spreadr function.

Let’s take a simple three-node network as example:

weighted_network <- matrix(
  c(0, 1, 9,
    1, 0, 0,
    9, 0, 0), nrow=3, byrow=TRUE)
colnames(weighted_network) <- c("a", "b", "c")
rownames(weighted_network) <- c("a", "b", "c")

# To visualise the network only --- this is not necessary for spreadr
weighted_igraph <- graph_from_adjacency_matrix(
  weighted_network, mode="undirected", weighted=TRUE)
plot(weighted_igraph, edge.width=E(weighted_igraph)$weight)

If we let the node "a" have initial activation = 10, we should expect to see 9 units of activation go to "c", and 1 unit to "b", for retention = 0.

spreadr(
  weighted_network, data.frame(node="a", activation=10),
  time=1, retention=0, include_t0=TRUE)
##   node activation time
## 1    a         10    0
## 2    b          0    0
## 3    c          0    0
## 4    a          0    1
## 5    b          1    1
## 6    c          9    1

Using a directed network

Simply pass a directed network (igraph object or adjacency matrix) to the spreadr function.

Let’s take a simple three-node network as example:

directed_network <- matrix(
  c(0, 1, 0,
    0, 0, 1,
    0, 0, 0), nrow=3, byrow=TRUE)
colnames(directed_network) <- c("a", "b", "c")
rownames(directed_network) <- c("a", "b", "c")

# To visualise the network only --- this is not necessary for spreadr
directed_igraph <- graph_from_adjacency_matrix(
  directed_network, mode="directed")
plot(directed_igraph, edge.width=E(directed_igraph)$weight)

If we let the node "b" have initial activation = 10, we should expect to see all 10 units of activation go to "c", and none go to "a", for retention = 0.

spreadr(
  directed_network, data.frame(node="b", activation=10),
  time=1, retention=0, include_t0=TRUE)
##   node activation time
## 1    a          0    0
## 2    b         10    0
## 3    c          0    0
## 4    a          0    1
## 5    b          0    1
## 6    c         10    1

Repeating the simulation across different parameters

There are a few ways to do this, but here we will showcase a relatively simple method using data.frames.

Suppose we want to simulate spreading activation across four different parameter sets of (retention, decay): (0, 0), (0.5, 0), (0, 0.5), and (0.5, 0.5). We would first record down all those parameters which will differ into a data.frame:

params <- data.frame(
  retention=c(0, 0.5,   0, 0.5),
  decay=    c(0,   0, 0.5, 0.5))

Then, we prepare the common parameters that will not differ.

network <- matrix(
  c(0, 1,
    0, 0), nrow=2, byrow=TRUE)
start_run <- data.frame(node=1, activation=10)

Now, to run the simulate once for each set of different parameters, we simply apply over the rows of params.

apply(params, 1, function(row)
  spreadr(
    network, start_run,
    time=2, include_t0=TRUE,
    retention=row[1], decay=row[2]))
## [[1]]
##   node activation time
## 1    1         10    0
## 2    2          0    0
## 3    1          0    1
## 4    2         10    1
## 5    1          0    2
## 6    2         10    2
## 
## [[2]]
##   node activation time
## 1    1       10.0    0
## 2    2        0.0    0
## 3    1        5.0    1
## 4    2        5.0    1
## 5    1        2.5    2
## 6    2        7.5    2
## 
## [[3]]
##   node activation time
## 1    1       10.0    0
## 2    2        0.0    0
## 3    1        0.0    1
## 4    2        5.0    1
## 5    1        0.0    2
## 6    2        2.5    2
## 
## [[4]]
##   node activation time
## 1    1     10.000    0
## 2    2      0.000    0
## 3    1      2.500    1
## 4    2      2.500    1
## 5    1      0.625    2
## 6    2      1.875    2

Reporting bugs or issues

First, please check that you are on the development build. This is because the development build contains a more up-to-date version of spreadr, which may include various bug fixes, as compared to the stable CRAN build.

If your problems persist, please report them on our Github repository.

Session Info

sessionInfo()
## R version 4.0.1 (2020-06-06)
## Platform: x86_64-apple-darwin17.0 (64-bit)
## Running under: macOS Catalina 10.15.7
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRblas.dylib
## LAPACK: /Library/Frameworks/R.framework/Versions/4.0/Resources/lib/libRlapack.dylib
## 
## locale:
## [1] C/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## attached base packages:
## [1] stats     graphics  grDevices utils     datasets  methods   base     
## 
## other attached packages:
## [1] ggplot2_3.3.3 igraph_1.2.6  spreadr_0.2.0 Rcpp_1.0.6   
## 
## loaded via a namespace (and not attached):
##  [1] compiler_4.0.1    pillar_1.5.1      highr_0.8         tools_4.0.1      
##  [5] digest_0.6.27     evaluate_0.14     lifecycle_1.0.0   tibble_3.1.0     
##  [9] gtable_0.3.0      lattice_0.20-41   debugme_1.1.0     pkgconfig_2.0.3  
## [13] rlang_0.4.10      Matrix_1.3-2      DBI_1.1.1         yaml_2.2.1       
## [17] xfun_0.21         withr_2.4.1       stringr_1.4.0     dplyr_1.0.4      
## [21] knitr_1.31        generics_0.1.0    vctrs_0.3.6       tidyselect_1.1.0 
## [25] grid_4.0.1        glue_1.4.2        R6_2.5.0          fansi_0.4.2      
## [29] rmarkdown_2.6     farver_2.0.3      purrr_0.3.4       magrittr_2.0.1   
## [33] scales_1.1.1      htmltools_0.5.1.1 ellipsis_0.3.1    assertthat_0.2.1 
## [37] colorspace_2.0-0  labeling_0.4.2    utf8_1.1.4        stringi_1.5.3    
## [41] munsell_0.5.0     crayon_1.4.1.9000

References

Chan, Kit Ying, and Michael S. Vitevitch. 2009. “The Influence of the Phonological Neighborhood Clustering-Coefficient on Spoken Word Recognition.” Journal of Experimental Psychology. Human Perception and Performance 35 (6): 1934–49. https://doi.org/10.1037/a0016902.

Siew, Cynthia S. Q. 2019. “Spreadr: An R Package to Simulate Spreading Activation in a Network.” Behavior Research Methods 51 (2): 910–29. https://doi.org/10.3758/s13428-018-1186-5.

Vitevitch, Michael S., Gunes Ercal, and Bhargav Adagarla. 2011. “Simulating Retrieval from a Highly Clustered Network: Implications for Spoken Word Recognition.” Frontiers in Psychology 2. https://doi.org/10.3389/fpsyg.2011.00369.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.