Skip to content

Latest commit

 

History

History
95 lines (72 loc) · 3.19 KB

README.md

File metadata and controls

95 lines (72 loc) · 3.19 KB

contrast

contrast is a Python implementation of the STUCCO algorithm (URL) that allows the ability to learn association rules with significant enrichment in one group over another. As a result, the produced association rules can help "describe" indicators aligned to a group of interest.

Dependencies

  • Python 3.7+
  • numpy (1.15+)
  • pandas (0.23+)

Such libraries can also be installed using the Anaconda Python installer.

Examples

A pandas DataFrame is needed to drive contrast-set learning. For demonstration purposes, we shall leverage existing seaborn datasets.

# read-in data and produce a DataFrame
from seaborn import load_dataset
frame = load_dataset('diamonds')

To execute the algorithm, we leverage the ContrastSetLearner class. A minimally working instance needs two arguments:

  • A pandas DataFrame
  • A feature name, group_feature
# a skeleton object; no contrast-set analysis has taken place.
from stucco import ContrastSetLearner
learner = ContrastSetLearner(frame, group_feature='color')

A feature can have many states, i.e. color = {'D', 'E', 'F', 'G'}. Thus, the goal of contrast-set learning is to gauge what rules are enriched across different group states. To make this happen, we leverage learner.learn() to enumerate rule abundance.

Considerations

We recommend tweaking learner.learn() parameter values, namely max_length. This parameter dictates the maximum rule length following derivation of its canonical combinations. For example: suppose we have the rule ['a = 1', 'b = 2', 'c = 3']. If max_length=2, all rule combinations, of length max_length, are produced. Due to this combinatorial function, an important consideration must be made:

  • Large max_length: increases runtime, possibility of intelligible rules.
  • Small max_length: quick runtime, few intelligible rules.
# derive 3-length combinations of a rule and enumerate their abundance.
learner.learn(max_length=3)

To determine if a rule, A, is enriched in a desired groups' state, B, rule counts get modeled as a 2 x 2 contingency matrix, m. We denote "not" symbol as ~:

m B ~B
A p(A, B) p(A, ~B)
~A p(~A, B) p(~A, ~B)

Given m, we can quantify rule abundance using several statistical metrics:

  • Support: p(A, B)
  • Lift: p(A, B) / p(A) * p(B)
  • Confidence: max(p(A, B) / p(A), p(A, B) / p(B))

These metrics are invoked in learner.score(). Their collective outputs must exceed user-provided thresholds in order to be deemed enriched in a state. Following completion of quantification, output, a DataFrame that references the rule, its group state, and its satisfactory lift score, is returned.

output = learner.score(min_lift=3)

Considerations

Increasing learner.score() parameter arguments renders scoring to be more stringent, and thus returns fewer intelligible rules. Therefore, we recommend experimenting with such parameter-values, namely min_lift and min_support.