CODEWITHSUNDEEP
rigrapharulesvisnetwork
yc.koong
yc.koong

Reputation: 175

r convert igraph into visNetwork

I found a way to convert igraph into visNetwork (refer to Interactive arules with arulesViz and visNetwork). Suppose before and after conversion from igraph to visNetwork should be the same, but my result shows after convert into visNetwork, the results are different.

I'll try demonstrate the issue using sample data data("Groceries") from Library(arules).

#Pre-defined library
library(arules)
library(arulesViz)
library(visNetwork)
library(igraph)

#Get sample data & get association rules
data("Groceries")
rules <- apriori(Groceries, parameter=list(support=0.01, confidence=0.4))
rules <- head(sort(rules, by="lift"), 10)

#Convert rules to data.table
library(data.table)
rules_dt <- data.table( lhs = labels( lhs(rules) ), 
                        rhs = labels( rhs(rules) ), 
                        quality(rules) )[ order(-lift), ]

print all rules in table format (sort by lift)

enter image description here

Plot top 10 association rules via using igraph

ig <- plot(rules, method="graph", control=list(type="items"))

enter image description here

Note: Based on association rules i plot a network diagram via using igraph, everything is correct. Next i'll try to convert existing igraph to visNetwork, then we compare the results.

tf <- tempfile( )
saveAsGraph(rules, file = tf, format = "dot" )
# clean up temp file if desired
#unlink(tf)

# Convert igraph to dataframe
ig_df <- as_data_frame(ig, what = "both")

# Plot visNetwork
visNetwork(
  nodes = data.frame(
     id = ig_df$vertices$name
 ,value = ig_df$vertices$lift # could change to lift or confidence
 ,title = ifelse(ig_df$vertices$label == "",ig_df$vertices$name, 
 ig_df$vertices$label)
 ,ig_df$vertices
 ), 
edges = ig_df$edges
) %>%
visEdges(arrows ="to") %>%  
visOptions( highlightNearest = T )

Plot top 10 association rules via using visNetwork enter image description here

Note: For visNetwork diagram, the size of intercept node indicate "lift", the higher the lift, the larger the size of intercept node; unlike igraph diagram, size of intercept node indicate "support", while colour of intercept node indicate "lift".

Let's compare the igraph and visNetwork

enter image description here By referring to the association rules in table format, rules no.10 (rules with smallest "lift"), suppose the size of the intercept node is the smallest, but end up it's not the smallest.

Problems

I tried to further drill down into ig_df <- get.data.frame( ig, what = "both" ), and i found something weird on ig_df$vertices table, which generated from as_data_frame function from library(igraph). enter image description here I found that the assoc10 (intercept node for association rules no.10) actually had NA for all variables (i.e. "support", "confidence", "lift" & "count"), more precisely the dimension for columns "support", "confidence", "lift" & "count" in ig_df$vertices are moving up one row! Kindly correct me if i'm wrong..

Conclusion Since the key to convert an igraph into visNetwork is to use this as_data_frame to get extract all data from an igraph and convert those data into dataframe, then plot visNetwork using the data from extracted dataframe. But due to extract issue when using as_data_frame to extract data from igraph, so the result from also different.

Question: is this a bug? or i made a mistake on my code? Any suggestion is welcome. Thanks!

Upvotes: 5

Views: 1050

Answers (1)

eporetsky
eporetsky

Reputation: 11

A year later... but I already typed everything so might as well.

If I understand it correctly, this is the source of your troubles -

value = ig_df$vertices$lift # could change to lift or confidence

The easiest solution is to assign your lift values to size because in visNetwork size: Number. Default to 25. The size is used to determine the size of node shapes that do not have the label inside of them. These shapes are: image, circularImage, diamond, dot, star, triangle, triangleDown, square and icon. I'm not sure how values works here but I think you could use values if you also provide appropriate scaling. https://www.rdocumentation.org/packages/visNetwork/versions/2.0.9/topics/visNodes

So the solution is:

size = ig_df$vertices$lift # could change to lift or confidence

Note that the nodes that have NA lift are set to a size 25 by default and that with a size 2 or 3 lift you can barely see the nodes. You could raise the lift size exponentially which will increase the size of the nodes and exaggerate the differences in lift.

ig_df <- as_data_frame(ig, what = "both")
ig_df$vertices["lift"] <- ig_df$vertices["lift"] ** 3

And I couldn't reproduce your shifty table problem, mine looked fine, hopefully it solves itself!

Upvotes: 1

Related Questions