BiRankFr: BiRank algorithm

In BavoDC/BiRankFraud: Computing fraud score using the BiRank algorithm

BiRank algorithm

Description

BiRank algorithm to compute fraud scores.

Usage

BiRankFr(
  sNetwork,
  fraudMat,
  Today = Sys.Date(),
  decayR = 0,
  decayF = 0,
  alpha = 0.85,
  maxiter = 1000,
  Epsilon = 1e-14,
  PrintProgress = F,
  pInit = NULL,
  cInit = NULL,
  ConvCriterion = c("Sep", "Whole", "Order")
)

Arguments

sNetwork	Data frame containing the variables startNode, endNode and possibly Date.
fraudMat	Data frame containing the variables FraudInd (binary variable indicatinf fraudulent claimes) and possibly Date.
Today	Date of analysis (e.g. 1/1/2020), default is Sys.Date(). Supply either as object with class Date or as character string in the format %d/%m/%Y.
decayR	Parameter for exponential decay of recency of relation (in weeks).
decayF	Parameter for exponential decay of recency of fraud (in weeks).
alpha	Damping factor for propagation algorithm (return to start).
maxiter	Maximum number of iterations for propagation algorithm.
Epsilon	Positive convergence tolerance \epsilon.
PrintProgress	Logical, indicates whether progress of the algorithm has to be printed.
pInit	Initial values for the party score vector p.
cInit	Initial values for the fraud score vector c.
ConvCriterion	Which convergence criterion to use. "Sep" uses ||p - p_{old}||_{2} / ||p_{old}||_{2} < \epsilon and ||c - c_{old}||_{2} / ||c_{old}||_{2} < \epsilon. "Whole" uses ||x - x_{old}||_{2} / ||x_{old}||_{2} < \epsilon with x = c(c, p). "Order" uses the same convergence criterion as "Sep" and checks if the order of the elements of \bold{p} and \bold{c} has not changed since the previous iteration.

Value

A list with the following components:

ResultsClaims	A data frame containing the claim IDs, fraud scores, scaled and normalized fraud scores.
ResultsParties	A data frame containing the party IDs, party scores, scaled and normalized party scores.
AdjacencyMatrix	Adjancency/weight matrix indicating which nodes are connected.
iter	Number of iterations that the algorithm needed to converge.
Converged	Logical, indicating whether the algorithm converged.

Examples

library(BiRankFraud)
NetwLabel = data.frame(
startNode = c('P2', 'P3', 'P3', 'C1', 'C1', 'C5', 'P1', 'P4', 'C2', 'C2'),
endNode = c('C3', 'C3', 'C4', 'P2', 'P3', 'P3', 'C1', 'C5', 'P1', 'P4'),
stringsAsFactors = FALSE
)
NetwLabel[grepl("C", NetwLabel$startNode), 1:2] = NetwLabel[grepl("C", NetwLabel$startNode), 2:1]
NetwLabel = NetwLabel[order(NetwLabel$startNode), ]
NetwLabel$startNode = as.numeric(gsub("P", "", NetwLabel$startNode))
NetwLabel$endNode   = as.numeric(gsub("C", "", NetwLabel$endNode))
NetwLabel

c0      = c(rep(0, 3), 1, 0)
Results = BiRankFr(NetwLabel, data.frame(FraudInd = c0))
Results

BavoDC/BiRankFraud documentation built on Aug. 30, 2023, 5:13 a.m.