Elicitation

pypbl.elicitation

Core module for preference based learning.

class pypbl.elicitation.BayesPreference(data)

Class for preference based learning using Bayesian inference

Parameters:data (object) – Pandas DataFrame with columns as features and index as item names
add_indifferent_preference(a, b)

Adds a preference based on a == b

Parameters:
  • a (str) – name of item a that has equal preference to item b
  • b (str) – name of item b
Raises:

ValueError

add_strict_preference(a, b)

Adds a preference based on a > b

Parameters:
  • a (str) – name of item a that is preferred over item b
  • b (str) – name of item b
Raises:

ValueError

compute_entropy(x)

Compute entropy of a new preference

Parameters:x (list) – pair of items
Returns:Entropy for the pair of items
indifferent_log_probability(preference, weights)

Computes the log probability for an indifferent preference

Parameters:
  • preference (tuple) – indifference preference relationship (a == b)
  • weights (Numpy Array) – weights for each feature
Returns:

Log probability of indifferent preference given weights

infer_weights(method='MAP', iterations=100)

Infer weights for each attribute based on preferences. method=’MAP’ uses gradient based optimisation to compute the maximum a posteriori, method=’mean’ uses sampling to compute a better estimate of the weights when using non-normal priors

Parameters:
  • method (str) – method used for inference
  • iterations (int) – number of iterations to use when method=’mean’
Returns:

Estimated weights using Bayesian inference

Raises:
  • AttributeError
  • ValueError
log_probability(weights)

Computes the log posterior probability based on the sum of each strict and indifferent preference probability

Parameters:weights (Numpy Array) – weights for each feature
Returns:Log posterior probability given preference data and weights
rank()

Rank items based on the inferred weights

Returns:Pandas DataFrame of ordered items and utility values
remove_last_strict_preference()

Removed the most recently added strict preference

set_priors(priors)

Set priors

Parameters:priors (list) – List of priors
Raises:AttributeError
strict_log_probability(preference, weights)

Computes the log probability for a strict preference

Parameters:
  • preference (tuple) – strict preference relationship (a > b)
  • weights (Numpy Array) – weights for each feature
Returns:

Log probability of strict preference given weights

suggest(method='random')

Suggest a new item to compare with the most preferred item method=’random’ chooses a new pair that includes the top ranked item at random, method=’max_variance’ chooses a pair that includes the top ranked item and the item with greatest uncertainty, method=’min_entropy’ chooses a pair that includes the top ranked item and minimises expected entropy

Parameters:method (str) – suggestion method one of ‘random’, ‘max_variance’ or ‘min_entropy’
Returns:Pair of items that includes the top ranked item and another item based on the selected method
Raises:ValueError
suggest_new_pair(method='random')

Suggest a new pair of items with minimum entropy method=’random’ chooses a new pair at random, method=’min_entropy’ chooses a new pair that minimises expected entropy

Parameters:method (str) – suggestion method one of ‘random’ or ‘min_entropy’
Returns:Pair of items that based on the selected method
Raises:ValueError