Dataset Level XAI

Variable Importance

Dashboard options

  • Maximum variables in Variable Importance
  • Display boxplots over Variable Importance
  • Left margin for variables names with values

Connector options

# Size of random subset. For large datasets can speedup calculations.
# Use None to select whole dataset.
arena.set_option('VariableImportance', 'N', None)
# Number of permutations.
arena.set_option('VariableImportance', 'B', 10)

Partial Dependence

Dashboard options

  • Left margin for variables values - for categorical variables

Connector options

# Method of selecting points: 'quantile' or 'uniform'
arena.set_option('PartialDependence', 'grid_type', 'quantile')
# Maximum number of points for profile. Final number can
# be lower using 'quantile' method.
arena.set_option('PartialDependence', 'grid_points', 101)
# Size of random subset. For large datasets can speedup calculations.
# Use None to select whole dataset.
arena.set_option('PartialDependence', 'N', 500)

Accumulated Dependence

Dashboard options

  • Left margin for variables values - for categorical variables

Connector options

# Method of selecting points: 'quantile' or 'uniform'
arena.set_option('AccumulatedDependence', 'grid_type', 'quantile')
# Maximum number of points for profile. Final number can
# be lower using 'quantile' method.
arena.set_option('AccumulatedDependence', 'grid_points', 101)
# Size of random subset. For large datasets can speedup calculations.
# Use None to select whole dataset.
arena.set_option('AccumulatedDependence', 'N', 500)

Shapley Values Dependence

Shapley Values were described in the page about observation level charts. In this chart we try to visualise how they depend on value of variable.

Dashboard options

  • Left margin for variables values - for categorical variables
  • Display error bars over Shapley dependence points - for numerical variables
  • Shapley dependence jitter range as per mill of chart range - for numerical variables

Connector options

# Size of random subset.
arena.set_option('DatasetShapleyValues', 'N', 500)
# Number of permutations for each observation.
# For best utilization thus should be multiplicity of parallel processes.
arena.set_option('DatasetShapleyValues', 'B', 4)
# Number of parallel processes
arena.set_option('DatasetShapleyValues', 'cpus', 4)

Shapley Values Variable Importance

Shapley Values can be used as a measure of importance. For this chart, we assume that important variables are those with large absolute contributions to the final prediction.

Dashboard options

  • Display boxplots over Variable Importance
  • Maximum variables in Variable Importance
  • Left margin for variables names with values

Connector options

# Size of random subset.
arena.set_option('DatasetShapleyValues', 'N', 500)
# Number of permutations for each observation.
# For best utilization thus should be multiplicity of parallel processes.
arena.set_option('DatasetShapleyValues', 'B', 4)
# Number of parallel processes
arena.set_option('DatasetShapleyValues', 'cpus', 4)
Observation Level XAI Fairness