def calc_collapsed_preds(indiv_predictions, subject_clusters, num_clusters):
collapsed = np.zeros((len(subject_clusters), len(indiv_predictions[0])))
for ans in range(len(indiv_predictions[0])):
indiv = np.array([ip[ans] for ip in indiv_predictions])
mns = clusters.compute_means(subject_clusters, indiv, num_clusters)
collapsed[:, ans] = [mns[subject_clusters[s]] for s in range(len(subject_clusters))]
return collapsed.tolist()
def calc_collapsed_param_bts(personal_probs, meta_params, meta_distribution_cdf, subject_clusters, answer_boundaries):
"""Make composite subjects and run bts on them.
Args:
personal_probs: list where each element is a subject's personal prob
meta_params: list where each element gives meta parameters induced from a subject. (A tuple of alpha, beta in beta distribution case.)
meta_distribution_cdf: function which calculates the CDF of the meta distribution
subject_clusters: np.array giving the cluster each subject is assigned to. Used to show which constructed answer a subject endorsed.
answer_boundaries: list where each element is a tuple giving (lower, upper) boundary of the answer. Answers needn't be in order.
Returns:
dict where each field (bts, surprise, accuracy) points to an np.array with an elt for each subject.
"""
num_subjects = len(subject_clusters)
num_clusters = len(answer_boundaries) #is this right?
cluster_personal = clusters.compute_means(subject_clusters, personal_probs, num_clusters)
collapsed_personal = [cluster_personal[subject_clusters[s]] for s in range(num_subjects)]
indiv_predictions = calc_predictions_cdf(meta_params, meta_distribution_cdf, answer_boundaries)
collapsed = calc_collapsed_preds(indiv_predictions, subject_clusters, num_clusters)
return calc_non_binary_bts(collapsed_personal, collapsed, subject_clusters, answer_boundaries)
