from . import MetricItem
import itertools
import numpy as np
import scipy.spatial.distance as distance
class Diversity(MetricItem):
def __init__(self, items_rated_by_user, k=10):
super(Diversity, self).__init__()
self.metric = CosineSimilarity(items_rated_by_user)
self.k = k
def evaluate(self, qid, targets, items=None):
diversity = 0.0
for items in itertools.combinations(items[:self.k], 2):
diversity += (1 - self.metric.similarity(items[0], items[1]))
return diversity / (self.k * (self.k - 1) * 0.5)
class CosineSimilarity:
def __init__(self, items_rated_by_user):
# Create items index and users index
self.items_index = {}
self.users_index = {}
current_item_id = 0
current_user_id = 0
for user, items_list in items_rated_by_user.items():
self.users_index[user] = current_user_id
current_user_id += 1
for item in items_list:
if item not in self.items_index:
self.items_index[item] = current_item_id
current_item_id += 1
# Create an item user matrix
self.matrix = np.full((current_item_id, current_user_id), 0)
for user, items_list in items_rated_by_user.items():
user_index = self.users_index[user]
for item in items_list:
item_index = self.items_index[item]
self.matrix[item_index, user_index] += 1
def similarity(self, item_i, item_j):
try:
index_i = self.items_index[item_i]
index_j = self.items_index[item_j]
except KeyError:
return 0
array_i = self.matrix[index_i]
array_j = self.matrix[index_j]
if array_i.sum() == 0 or array_j.sum() == 0:
return 0
else:
return 1 - distance.cosine(array_i, array_j)
