Skip to contents

Advanced lattice metrics

Source: vignettes/advanced_lattice_metrics.Rmd
advanced_lattice_metrics.Rmd

Introduction

Beyond the standard support (frequency), fcaR provides advanced metrics to analyze the quality and robustness of formal concepts, especially in fuzzy settings.

  • Stability: Measures the robustness of a concept against noise.
  • Separation: Measures the specificity of the objects covered by a concept.
  • Fuzzy Density: Measures the cohesion of the concept in the original data.

Stability

Intensional stability is defined as the probability that the intent of a concept is preserved when a random subset of its extent is removed.

σ(C)=|{A⊆Ext(C)∣A′=Int(C)}|2|Ext(C)| \sigma(C) = \frac{|\{ A \subseteq \text{Ext}(C) \mid A' = \text{Int}(C) \}|}{2^{|\text{Ext}(C)|}}

Let’s compute it for the planets dataset:

fc <- FormalContext$new(planets)
fc$find_concepts()

stab <- fc$concepts$stability()
head(stab)
#> [1] 1.000000 0.703125 0.750000 0.750000 0.750000 0.562500

Concepts with stability close to 1 are very robust (likely real patterns), while those near 0 are unstable (likely artifacts).

Separation

Separation measures how many objects are “unique” to a concept CC, i.e., they are covered by CC but not by any of its direct subconcepts (descendants).

Sep(C)=|Ext(C)|−|⋃K≺CExt(K)| \text{Sep}(C) = |\text{Ext}(C)| - |\bigcup_{K \prec C} \text{Ext}(K)| 

sep <- fc$concepts$separation()
head(sep)
#> [1] 1 0 2 2 2 0

A high separation indicates that the concept introduces a significant set of new objects into the hierarchy.

Fuzzy density

In Fuzzy FCA, concepts are “rectangles” of high values in the matrix, but not necessarily all 1s. Density measures the average value of the relation II within the concept.

ρ(C)=∑g∈Ext(C),m∈Int(C)I(g,m)|Ext(C)|⋅|Int(C)| \rho(C) = \frac{\sum_{g \in \text{Ext}(C), m \in \text{Int}(C)} I(g, m)}{|\text{Ext}(C)| \cdot |\text{Int}(C)|} 

# For binary data, density is always 1 (or 0 for empty concepts)
# We need to pass the original matrix I
dens <- fc$concepts$density(I = fc$incidence())
head(dens)
#> [1] 0 0 1 1 1 1

Putting it all together

We can combine these metrics to filter and analyze the lattice.

df <- data.frame(
  Concept = 1:fc$concepts$size(),
  Support = fc$concepts$support(),
  Stability = stab,
  Separation = sep,
  Density = dens
)

# Concepts with high stability and separation
subset(df, Stability > 0.8 & Separation > 0)
#>   Concept Support Stability Separation Density
#> 1       1       0         1          1       0