Given a large set of problems and their individual solutions case based reasoning seeks to solve a new problem by referring to the solution of that problem which is "most similar" to the new problem. Crucial in case based reasoning is the decision which problem "most closely" matches a given new problem. The basic idea is to define a family of distance functions and to use these distance functions as parameters of local averaging regression estimates of the final result. Then that distance function is chosen for which the resulting estimate is optimal with respect to a certain error measure used in regression estimation. The idea is based on: Dippon J. et al. (2002)
The R package case-based-reasoning provides an R interface case based reasoning using machine learning.
This R package provides two methods case based reasoning by using an endpoint:
Linear, logistic, and Cox regression
Proximity and Depth Measure extracted from a fitted random forest (ranger package)
Besides the functionality of searching similar cases, some additional features are included:
automatic validation of the key variables between the query and similar cases dataset
checking proportional hazard assumption for the Cox Model
C++-functions for distance calculation
In the first example, we use the Cox-Model and the
ovarian data set from the
survival package. In the first step we initialize the R6 data object.
library(tidyverse) library(survival) library(CaseBasedReasoning) ovarian$resid.ds <- factor(ovarian$resid.ds) ovarian$rx <- factor(ovarian$rx) ovarian$ecog.ps <- factor(ovarian$ecog.ps) # initialize R6 object coxBeta <- CoxBetaModel$new(Surv(futime, fustat) ~ age + resid.ds + rx + ecog.ps)
After the initialization, we may want to get for each case in the query data the most similar case from the learning data.
n <- nrow(ovarian)trainID <- sample(1:n, floor(0.8 * n), F)testID <- (1:n)[-trainID]# fit modelovarian[trainID, ] %>%coxBeta$fit()# get similar casesovarian[testID, ] %>%coxBeta$get_similar_cases(queryData = ovarian[testID, ], k = 3) -> matchedData
You may extract then the similar cases and the verum data and put them together:
Note 1: In the initialization step, we dropped all cases with missing values in the variables of
endPoint. So, you need to make sure that NA handling is done by you.
Note 2: The
data.table returned from
coxBeta$get_similar_cases has four additional columns:
caseId: By this column you may map the similar cases to cases in data, e.g. if you had chosen
k = 3, then the first three elements in the column
2and so on). This means that this three cases are the three most similar cases to case
0in verum data.
scDist: The calculated distance
scCaseId: Grouping number of query with matched data
group: Grouping matched or query data
Alternatively, you may just be interested in the distance matrix, then you go this way:
ovarian %>%coxBeta$calc_distance_matrix() -> distMatrix
coxBeta$calc_distance_matrix() calculates the full distance matrix. This matrix the dimension: cases of data versus cases of query data. If the query dataset is bot available, this functions calculates a n times n distance matrix of all pairs in data.
The distance matrix is saved internally in the cbrCoxModel object:
In the second example, we present the Random Forest model for a distance measure approximation applied on the
ovarian data set from the
survival package. This package offers two ways for distance/similarity calculation (see documentation):
Let's initialize the R6 data object.
library(tidyverse)library(survival)library(CaseBasedReasoning)ovarian$resid.ds <- factor(ovarian$resid.ds)ovarian$rx <- factor(ovarian$rx)ovarian$ecog.ps <- factor(ovarian$ecog.ps)# initialize R6 objectrfSC <- RFModel$new(Surv(futime, fustat) ~ age + resid.ds + rx + ecog.ps)
All cases with missing values in the learning and end point variables are dropped (
na.omit) and the reduced data set without missing values is saved internally. You get a text output on how many cases were dropped.
character variables will be transformed to
Optionally, you may want to adjust some parameters in the fitting step of the random forest algorithm. Possible arguments are: ,
splitrule. The documentation of this parameters can be found in the ranger R-package. Furthermore, you are able to choose the two distance measures:
Proximity: the proximity matrix
Depth(Default): Calculates the average edge length over all trees
This can be done by
rfSC$set_dist(distMethod = "Proximity")
All other steps (excluding checking for proportional hazard assumption are the same as for the Cox-Model).
n <- nrow(ovarian)trainID <- sample(1:n, floor(0.8 * n), F)testID <- (1:n)[-trainID]# fit modelovarian[trainID, ] %>%rfSC$fit()# get similar casesovarian[trainID, ] %>%rfSC$get_similar_cases(queryData = ovarian[testID, ], k = 3) -> matchedData
Distance Matrix Calculation:
ovarian %>%rfSC$calc_distance_matrix() -> distMatrix
PD Dr. Jürgen Dippon, Institut für Stochastik und Anwendungen, Universität Stuttgart
Dr. Simon Müller, TTI GmbH - MUON-STAT
Dr. Peter Fritz
Professor Dr. Friedel
The work was funded by the Robert Bosch Foundation. Special thanks go to Professor Dr. Friedel (Thoraxchirugie - Klinik Schillerhöhe).
Dippon et al. A statistical approach to case based reasoning, with application to breast cancer data (2002),
Friedel et al. Postoperative Survival of Lung Cancer Patients: Are There Predictors beyond TNM? (2012).
Englund and Verikas A novel approach to estimate proximity in a random forest: An exploratory study
Stuart, E. et al. Matching methods for causal inference: Designing observational studies