We (Edwin de Jonge and me) have recently updated our editrules package. The most important new features include (beta) support for categorical data. However, in this
post I'm going to show some visualizations we included, made possible by Gabor Csardi's awesome igraph package.
Make sure you run
before trying the code below.
First, let's load editrules' built-in editset:
> data(edits) > edits name edit description 1 b1 t == ct + p total balance 2 b2 ct == ch + cp cost balance 3 s1 p <= 0.6*t profit sanity 4 s2 cp <= 0.3*t personnel cost sanity 5 s3 ch <= 0.3*t housing cost sanity 6 p1 t >0 turnover positivity 7 p2 ch > 0 housing cost positivity 8 p3 cp > 0 personnel cost positivity 9 p4 ct > 0 total cost positivity
Here, edits is a data.frame with a "name" column, naming the rules, an "edit" column, with a character representation of the edit rules and a "description" column. The variables have the following meaning: t: turnover, ct total cost, p profit, ch housing costs and cp personnel cost. The rules demand balance accounts to add up (e.g.cost + profit equals turnover) and demand some sanity checks (e.g. profit can not exceed 60% of turnover).
The sanity checks here are completely fictional. To do anything useful with these rules, turn them into an editmatrix.
> (E <- editmatrix(edits)) Edit matrix: ct p t ch cp Ops CONSTANT b1 -1 -1 1.0 0 0 == 0 b2 1 0 0.0 -1 -1 == 0 s1 0 1 -0.6 0 0 <= 0 s2 0 0 -0.3 0 1 <= 0 s3 0 0 -0.3 1 0 <= 0 p1 0 0 -1.0 0 0 < 0 p2 0 0 0.0 -1 0 < 0 p3 0 0 0.0 0 -1 < 0 p4 -1 0 0.0 0 0 < 0 Edit rules: b1 : t == ct + p [ total balance ] b2 : ct == ch + cp [ cost balance ] s1 : p <= 0.6*t [ profit sanity ] s2 : cp <= 0.3*t [ personnel cost sanity ] s3 : ch <= 0.3*t [ housing cost sanity ] p1 : 0 < t [ turnover positivity ] p2 : 0 < ch [ housing cost positivity ] p3 : 0 < cp [ personnel cost positivity ] p4 : 0 < ct [ total cost positivity ]
Although the matrix representation and the textual representation have their merits, it is hard to see which rules are (indirectly) related via shared variables. This may be visualized by plotting the rules in a graph, where each variable and edit is a node, and a variable node is connected with an editnode if the variable occurs in the edits. Just do
The round, blue nodes represent variables and the square nodes represent edit rules.
You can see at a glance that everything is connected, so the editmatrix does not block into submatrices. If you want to leave out the variables, and just see how the edits are connected, use
Which you can try for yourself. We can do cooler stuff. For example, lets define a faulty record and detect which rules it violates.
> r <- c(ct = 100, ch = 30, cp = 70, p=30,t=130 ) > (v <- violatedEdits(E,r)) edit record b1 b2 s1 s2 s3 p1 p2 p3 p4 1 FALSE FALSE FALSE TRUE FALSE FALSE FALSE FALSE FALSE
So only edit s2 is violated. We can visualize this as well.
The complexity of error localization shown in a glance. We can try to adapt cp, but this might yield violation of b2 or p3. So here's the central question in error localization: what is the least (weighted) number of variables we have to change such that all violated rules can be obeyed without causing new violations? Editrules was actually written to answer this question. There are several functions performing this task, but here we'll use the low-level errorLocalizer function and plot the result.
adapt <- errorLocalizer(E,r)$searchBest()$adapt plot(E, violated=v, adapt=adapt )
So, in order to repair the record, the turnover needs to be altered and to make sure no other rules are violated, the profit p has to be altered as well.
If you don't like the colors or want to play with the igraph objects yourself, see the as.igraph or adjacency functions.
Oh, and if you wander which are the possible values to use for p and t, just substitute all the other values in the editmatrix:
> substValue(E,names(r)[!adapt],r[!adapt]) Edit matrix: ct p t ch cp Ops CONSTANT b1 0 -1 1.0 0 0 == 100 s1 0 1 -0.6 0 0 <= 0 s2 0 0 -0.3 0 0 <= -70 s3 0 0 -0.3 0 0 <= -30 p1 0 0 -1.0 0 0 < 0 Edit rules: b1 : t == p + 100 s1 : p <= 0.6*t s2 : 70 <= 0.3*t s3 : 30 <= 0.3*t p1 : 0 < t
The solution set to the above system of equations is the set of possible values for t and p.