R/getCnetFeatures.R

In rgmantovani/MfeatAnalysis: Extract meta-features from classification datasets

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

# R Code
# Network Metrics
# L. P. Garcia, A. C.P.L.F. Carvalho, A. C. Lorena 2014
# Measures based on complexity network analysis

# Please cite:
# Luís P.F. Garcia, André C.P.L.F. de Carvalho, Ana C. Lorena. Effect of label noise in the
# complexity of classification problems. Neurocomputing, v 160 (2015). 108 - 119.

# Requires: igraph R package

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

basic = function(graph) {
  edge = igraph::ecount(graph)
  rate = mean(igraph::degree(graph))
  aux  = c(edge, rate)
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

density = function(graph) {
  aux = igraph::graph.density(graph)
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

componets = function(graph) {
  tmp = igraph::clusters(graph, mode = "weak")
  aux = max(tmp$csize)
  return(aux)        
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

centrality = function(graph) {
  aux = mean(igraph::closeness(graph))
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

between = function(graph) {
  aux = mean(igraph::betweenness(graph))
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

hub = function(graph) {
  aux = mean(igraph::hub.score(graph)$vector)
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

giantComponent = function(graph) {
  cls = igraph::clusters(graph)
  aux = igraph::induced.subgraph(graph, which(cls$membership == which.max(cls$csize)))
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

clusterCof = function(graph) {
  g = giantComponent(graph)
  disthist = igraph::path.length.hist(g, directed = FALSE)$res

  if(length(disthist) == 0) {
    avg = 0
  } else {
    avg = weighted.mean(1:length(disthist), disthist)
  }

  cof = igraph::transitivity(graph, type = "undirected")
  aux = c(cof, avg)
  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

ennSup = function(data, epson) {
  
  dst = as.matrix(dist(data[,-ncol(data)], method = "minkowski"))
  aux = matrix(0, nrow(data), nrow(data), dimnames = list(rownames(data), rownames(data)))

  for(i in 1:nrow(aux)) {
    tmp1 = names(sort(dst[i,])[1:(epson*nrow(data))+1])
    tmp2 = rownames(data[data$Class == data[i,]$Class,])
    tmp3 = intersect(tmp1, tmp2)
    aux[i, tmp3] = 1
  }

  return(aux)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------

getCnetFeatures = function(data, epson = 0.15) {

  aux = ennSup(data, epson)
  graph = igraph::graph.adjacency(adjmatrix = aux, mode = "undirected")
  tmp = c(basic(graph), density(graph), componets(graph), centrality(graph), 
    between(graph), hub(graph), clusterCof(graph))

  names(tmp) = c("edges", "degree", "density", "maxComp", "closeness", "betweenness", 
    "clsCoef", "hubs", "avgPath")

  return(tmp)
}

# -------------------------------------------------------------------------------------------------
# -------------------------------------------------------------------------------------------------