def fit(self, item_weights, URM_train, selectTopK=False):
self.URM_train = check_matrix(URM_train, format='csc')
if self.URM_train.shape[1] != item_weights.shape[0]:
raise ValueError(
"ItemKNNCustomSimilarityRecommender: URM_train and item_weights matrices are not consistent. "
"The number of columns in URM_train must be equal to the rows in item_weights."
"Current shapes are: URM_train {}, item_weights {}".format(
self.URM_train.shape, item_weights.shape))
if item_weights.shape[0] != item_weights.shape[1]:
raise ValueError(
"ItemKNNCustomSimilarityRecommender: item_weights matrice is not square. "
"Current shape is {}".format(item_weights.shape))
# If no topK selection is required, just save the similarity
if not selectTopK:
if isinstance(item_weights, np.ndarray):
self.W = item_weights
self.sparse_weights = False
else:
self.W_sparse = check_matrix(item_weights, format='csr')
self.sparse_weights = True
return
else:
self.W_sparse = similarityMatrixTopK(item_weights,
forceSparseOutput=True,
k=self.topK)
self.sparse_weights = True
def applyPearsonCorrelation(self):
"""
Remove from every data point the average for the corresponding column
:return:
"""
self.dataMatrix = check_matrix(self.dataMatrix, 'csc')
interactionsPerCol = np.diff(self.dataMatrix.indptr)
nonzeroCols = interactionsPerCol > 0
sumPerCol = np.asarray(self.dataMatrix.sum(axis=0)).ravel()
colAverage = np.zeros_like(sumPerCol)
colAverage[nonzeroCols] = sumPerCol[nonzeroCols] / interactionsPerCol[
nonzeroCols]
# Split in blocks to avoid duplicating the whole data structure
start_col = 0
end_col = 0
blockSize = 1000
while end_col < self.n_columns:
end_col = min(self.n_columns, end_col + blockSize)
self.dataMatrix.data[self.dataMatrix.indptr[start_col]:self.dataMatrix.indptr[end_col]] -= \
np.repeat(colAverage[start_col:end_col], interactionsPerCol[start_col:end_col])
start_col += blockSize
def applyAdjustedCosine(self):
"""
Remove from every data point the average for the corresponding row
:return:
"""
self.dataMatrix = check_matrix(self.dataMatrix, 'csr')
interactionsPerRow = np.diff(self.dataMatrix.indptr)
nonzeroRows = interactionsPerRow > 0
sumPerRow = np.asarray(self.dataMatrix.sum(axis=1)).ravel()
rowAverage = np.zeros_like(sumPerRow)
rowAverage[nonzeroRows] = sumPerRow[nonzeroRows] / interactionsPerRow[
nonzeroRows]
# Split in blocks to avoid duplicating the whole data structure
start_row = 0
end_row = 0
blockSize = 1000
while end_row < self.n_rows:
end_row = min(self.n_rows, end_row + blockSize)
self.dataMatrix.data[self.dataMatrix.indptr[start_row]:self.dataMatrix.indptr[end_row]] -= \
np.repeat(rowAverage[start_row:end_row], interactionsPerRow[start_row:end_row])
start_row += blockSize
def __init__(self, URM_train):
super(Random, self).__init__()
# convert to csc matrix for faster column-wise sum
self.URM_train = check_matrix(URM_train, 'csr', dtype=np.float32)
self.compute_item_score = self.compute_score_random
def __init__(self, URM_train):
super(GlobalEffects, self).__init__()
self.URM_train = check_matrix(URM_train, 'csc', dtype=np.float32)
self.compute_item_score = self.compute_score_global_effects
def __init__(self, URM_recommendations_items):
super(PredefinedListRecommender, self).__init__()
# convert to csc matrix for faster column-wise sum
self.URM_recommendations = check_matrix(URM_recommendations_items, 'csr', dtype=np.int)
self.URM_train = sps.csr_matrix((self.URM_recommendations.shape))
def __init__(self, URM_train):
super(TopPop, self).__init__()
# convert to csc matrix for faster column-wise sum
self.URM_train = check_matrix(URM_train, 'csc', dtype=np.float32)
self.URM_train.eliminate_zeros()
self.compute_item_score = self.compute_score_top_pop
def __init__(self, ICM, URM_train, sparse_weights=True):
super(ItemKNNCBFRecommender, self).__init__()
self.ICM = ICM.copy()
# CSR is faster during evaluation
self.URM_train = check_matrix(URM_train.copy(), 'csr')
self.sparse_weights = sparse_weights
def fit(self):
# This command returns a numpy.matrix of size (1, nitems)
# Use np.ediff1d and NOT a sum done over the rows as there might be values other than 0/1
self.item_pop = np.ediff1d(self.URM_train.indptr)
self.URM_train = check_matrix(self.URM_train, 'csr', dtype=np.float32)
self.item_pop = np.asarray(self.item_pop).squeeze(
) # necessary to convert it into a numpy.array of size (nitems,)
def __init__(self,
URM_train,
ICM,
S_matrix_target,
recompile_cython=False):
super(CFW_D_Similarity_Cython, self).__init__()
if (URM_train.shape[1] != ICM.shape[0]):
raise ValueError(
"Number of items not consistent. URM contains {} but ICM contains {}"
.format(URM_train.shape[1], ICM.shape[0]))
if (S_matrix_target.shape[0] != S_matrix_target.shape[1]):
raise ValueError(
"Items imilarity matrix is not square: rows are {}, columns are {}"
.format(S_matrix_target.shape[0], S_matrix_target.shape[1]))
if (S_matrix_target.shape[0] != ICM.shape[0]):
raise ValueError(
"Number of items not consistent. S_matrix contains {} but ICM contains {}"
.format(S_matrix_target.shape[0], ICM.shape[0]))
self.URM_train = check_matrix(URM_train, 'csr')
self.S_matrix_target = check_matrix(S_matrix_target, 'csr')
self.ICM = check_matrix(ICM, 'csr')
self.n_items = self.URM_train.shape[1]
self.n_users = self.URM_train.shape[0]
self.n_features = self.ICM.shape[1]
self.sparse_weights = True
if recompile_cython:
print("Compiling in Cython")
self.runCompilationScript()
print("Compilation Complete")
def fit(self, lambda_user=10, lambda_item=25):
self.lambda_user = lambda_user
self.lambda_item = lambda_item
# convert to csc matrix for faster column-wise sum
# 1) global average
self.mu = self.URM_train.data.sum(
dtype=np.float32) / self.URM_train.data.shape[0]
# 2) item average bias
# compute the number of non-zero elements for each column
col_nnz = np.diff(self.URM_train.indptr)
# it is equivalent to:
# col_nnz = X.indptr[1:] - X.indptr[:-1]
# and it is **much faster** than
# col_nnz = (X != 0).sum(axis=0)
URM_train_unbiased = self.URM_train.copy()
URM_train_unbiased.data -= self.mu
self.item_bias = URM_train_unbiased.sum(axis=0) / (col_nnz +
self.lambda_item)
self.item_bias = np.asarray(self.item_bias).ravel(
) # converts 2-d matrix to 1-d array without anycopy
# 3) user average bias
# NOTE: the user bias is *useless* for the sake of ranking items. We just show it here for educational purposes.
# first subtract the item biases from each column
# then repeat each element of the item bias vector a number of times equal to col_nnz
# and subtract it from the data vector
URM_train_unbiased.data -= np.repeat(self.item_bias, col_nnz)
# now convert the csc matrix to csr for efficient row-wise computation
URM_train_unbiased_csr = URM_train_unbiased.tocsr()
row_nnz = np.diff(URM_train_unbiased_csr.indptr)
# finally, let's compute the bias
self.user_bias = URM_train_unbiased_csr.sum(
axis=1).ravel() / (row_nnz + self.lambda_user)
# 4) precompute the item ranking by using the item bias only
# the global average and user bias won't change the ranking, so there is no need to use them
#self.item_ranking = np.argsort(self.bi)[::-1]
self.URM_train = check_matrix(self.URM_train, 'csr', dtype=np.float32)
def compute_similarity(self, start_col=None, end_col=None, block_size=100):
"""
Compute the similarity for the given dataset
:param self:
:param start_col: column to begin with
:param end_col: column to stop before, end_col is excluded
:return:
"""
values = []
rows = []
cols = []
start_time = time.time()
start_time_print_batch = start_time
processedItems = 0
if self.adjusted_cosine:
self.applyAdjustedCosine()
elif self.pearson_correlation:
self.applyPearsonCorrelation()
elif self.tanimoto_coefficient or self.dice_coefficient or self.tversky_coefficient:
self.useOnlyBooleanInteractions()
# We explore the matrix column-wise
self.dataMatrix = check_matrix(self.dataMatrix, 'csc')
# Compute sum of squared values to be used in normalization
sumOfSquared = np.array(self.dataMatrix.power(2).sum(axis=0)).ravel()
# Tanimoto does not require the square root to be applied
if not (self.tanimoto_coefficient or self.dice_coefficient
or self.tversky_coefficient):
sumOfSquared = np.sqrt(sumOfSquared)
if self.asymmetric_cosine:
sumOfSquared_to_1_minus_alpha = np.power(
sumOfSquared, 2 * (1 - self.asymmetric_alpha))
sumOfSquared_to_alpha = np.power(sumOfSquared,
2 * self.asymmetric_alpha)
self.dataMatrix = check_matrix(self.dataMatrix, 'csc')
start_col_local = 0
end_col_local = self.n_columns
if start_col is not None and start_col > 0 and start_col < self.n_columns:
start_col_local = start_col
if end_col is not None and end_col > start_col_local and end_col < self.n_columns:
end_col_local = end_col
start_col_block = start_col_local
this_block_size = 0
# Compute all similarities for each item using vectorization
while start_col_block < end_col_local:
# Add previous block size
processedItems += this_block_size
end_col_block = min(start_col_block + block_size, end_col_local)
this_block_size = end_col_block - start_col_block
if time.time(
) - start_time_print_batch >= 30 or end_col_block == end_col_local:
columnPerSec = processedItems / (time.time() - start_time)
print(
"Similarity column {} ( {:2.0f} % ), {:.2f} column/sec, elapsed time {:.2f} min"
.format(
processedItems, processedItems /
(end_col_local - start_col_local) * 100, columnPerSec,
(time.time() - start_time) / 60))
sys.stdout.flush()
sys.stderr.flush()
start_time_print_batch = time.time()
# All data points for a given item
item_data = self.dataMatrix[:, start_col_block:end_col_block]
item_data = item_data.toarray().squeeze()
if self.use_row_weights:
#item_data = np.multiply(item_data, self.row_weights)
#item_data = item_data.T.dot(self.row_weights_diag).T
this_block_weights = self.dataMatrix_weighted.T.dot(item_data)
else:
# Compute item similarities
this_block_weights = self.dataMatrix.T.dot(item_data)
for col_index_in_block in range(this_block_size):
if this_block_size == 1:
this_column_weights = this_block_weights
else:
this_column_weights = this_block_weights[:,
col_index_in_block]
columnIndex = col_index_in_block + start_col_block
this_column_weights[columnIndex] = 0.0
# Apply normalization and shrinkage, ensure denominator != 0
if self.normalize:
if self.asymmetric_cosine:
denominator = sumOfSquared_to_alpha[
columnIndex] * sumOfSquared_to_1_minus_alpha + self.shrink + 1e-6
else:
denominator = sumOfSquared[
columnIndex] * sumOfSquared + self.shrink + 1e-6
this_column_weights = np.multiply(this_column_weights,
1 / denominator)
# Apply the specific denominator for Tanimoto
elif self.tanimoto_coefficient:
denominator = sumOfSquared[
columnIndex] + sumOfSquared - this_column_weights + self.shrink + 1e-6
this_column_weights = np.multiply(this_column_weights,
1 / denominator)
elif self.dice_coefficient:
denominator = sumOfSquared[
columnIndex] + sumOfSquared + self.shrink + 1e-6
this_column_weights = np.multiply(this_column_weights,
1 / denominator)
elif self.tversky_coefficient:
denominator = this_column_weights + \
(sumOfSquared[columnIndex] - this_column_weights)*self.tversky_alpha + \
(sumOfSquared - this_column_weights)*self.tversky_beta + self.shrink + 1e-6
this_column_weights = np.multiply(this_column_weights,
1 / denominator)
# If no normalization or tanimoto is selected, apply only shrink
elif self.shrink != 0:
this_column_weights = this_column_weights / self.shrink
#this_column_weights = this_column_weights.toarray().ravel()
if self.TopK == 0:
self.W_dense[:, columnIndex] = this_column_weights
else:
# Sort indices and select TopK
# Sorting is done in three steps. Faster then plain np.argsort for higher number of items
# - Partition the data to extract the set of relevant items
# - Sort only the relevant items
# - Get the original item index
relevant_items_partition = (
-this_column_weights).argpartition(self.TopK -
1)[0:self.TopK]
relevant_items_partition_sorting = np.argsort(
-this_column_weights[relevant_items_partition])
top_k_idx = relevant_items_partition[
relevant_items_partition_sorting]
# Incrementally build sparse matrix, do not add zeros
notZerosMask = this_column_weights[top_k_idx] != 0.0
numNotZeros = np.sum(notZerosMask)
values.extend(this_column_weights[top_k_idx][notZerosMask])
rows.extend(top_k_idx[notZerosMask])
cols.extend(np.ones(numNotZeros) * columnIndex)
start_col_block += block_size
# End while on columns
if self.TopK == 0:
return self.W_dense
else:
W_sparse = sps.csr_matrix((values, (rows, cols)),
shape=(self.n_columns, self.n_columns),
dtype=np.float32)
return W_sparse
def generateTrainData_low_ram(self):
print(self.RECOMMENDER_NAME + ": Generating train data")
start_time_batch = time.time()
# Here is important only the structure
self.similarity = Compute_Similarity(self.ICM.T,
shrink=0,
topK=self.topK,
normalize=False)
S_matrix_contentKNN = self.similarity.compute_similarity()
S_matrix_contentKNN = check_matrix(S_matrix_contentKNN, "csr")
self.writeLog(
self.RECOMMENDER_NAME +
": Collaborative S density: {:.2E}, nonzero cells {}".format(
self.S_matrix_target.nnz /
self.S_matrix_target.shape[0]**2, self.S_matrix_target.nnz))
self.writeLog(
self.RECOMMENDER_NAME +
": Content S density: {:.2E}, nonzero cells {}".format(
S_matrix_contentKNN.nnz /
S_matrix_contentKNN.shape[0]**2, S_matrix_contentKNN.nnz))
if self.normalize_similarity:
# Compute sum of squared
sum_of_squared_features = np.array(
self.ICM.T.power(2).sum(axis=0)).ravel()
sum_of_squared_features = np.sqrt(sum_of_squared_features)
num_common_coordinates = 0
estimated_n_samples = int(S_matrix_contentKNN.nnz *
(1 + self.add_zeros_quota) * 1.2)
self.row_list = np.zeros(estimated_n_samples, dtype=np.int32)
self.col_list = np.zeros(estimated_n_samples, dtype=np.int32)
self.data_list = np.zeros(estimated_n_samples, dtype=np.float64)
num_samples = 0
for row_index in range(self.n_items):
start_pos_content = S_matrix_contentKNN.indptr[row_index]
end_pos_content = S_matrix_contentKNN.indptr[row_index + 1]
content_coordinates = S_matrix_contentKNN.indices[
start_pos_content:end_pos_content]
start_pos_target = self.S_matrix_target.indptr[row_index]
end_pos_target = self.S_matrix_target.indptr[row_index + 1]
target_coordinates = self.S_matrix_target.indices[
start_pos_target:end_pos_target]
# Chech whether the content coordinate is associated to a non zero target value
# If true, the content coordinate has a collaborative non-zero value
# if false, the content coordinate has a collaborative zero value
is_common = np.in1d(content_coordinates, target_coordinates)
num_common_in_current_row = is_common.sum()
num_common_coordinates += num_common_in_current_row
for index in range(len(is_common)):
if num_samples == estimated_n_samples:
dataBlock = 1000000
self.row_list = np.concatenate(
(self.row_list, np.zeros(dataBlock, dtype=np.int32)))
self.col_list = np.concatenate(
(self.col_list, np.zeros(dataBlock, dtype=np.int32)))
self.data_list = np.concatenate(
(self.data_list, np.zeros(dataBlock,
dtype=np.float64)))
if is_common[index]:
# If cell exists in target matrix, add its value
# Otherwise it will remain zero with a certain probability
col_index = content_coordinates[index]
self.row_list[num_samples] = row_index
self.col_list[num_samples] = col_index
new_data_value = self.S_matrix_target[row_index, col_index]
if self.normalize_similarity:
new_data_value *= sum_of_squared_features[
row_index] * sum_of_squared_features[col_index]
self.data_list[num_samples] = new_data_value
num_samples += 1
elif np.random.rand() <= self.add_zeros_quota:
col_index = content_coordinates[index]
self.row_list[num_samples] = row_index
self.col_list[num_samples] = col_index
self.data_list[num_samples] = 0.0
num_samples += 1
if time.time(
) - start_time_batch > 30 or num_samples == S_matrix_contentKNN.nnz * (
1 + self.add_zeros_quota):
print(self.RECOMMENDER_NAME +
": Generating train data. Sample {} ( {:.2f} %) ".format(
num_samples, num_samples / S_matrix_contentKNN.nnz *
(1 + self.add_zeros_quota) * 100))
sys.stdout.flush()
sys.stderr.flush()
start_time_batch = time.time()
self.writeLog(
self.RECOMMENDER_NAME +
": Content S structure has {} out of {} ( {:.2f}%) nonzero collaborative cells"
.format(num_common_coordinates, S_matrix_contentKNN.nnz,
num_common_coordinates / S_matrix_contentKNN.nnz * 100))
# Discard extra cells at the left of the array
self.row_list = self.row_list[:num_samples]
self.col_list = self.col_list[:num_samples]
self.data_list = self.data_list[:num_samples]
data_nnz = sum(np.array(self.data_list) != 0)
data_sum = sum(self.data_list)
collaborative_nnz = self.S_matrix_target.nnz
collaborative_sum = sum(self.S_matrix_target.data)
self.writeLog(
self.RECOMMENDER_NAME +
": Nonzero collaborative cell sum is: {:.2E}, average is: {:.2E}, "
"average over all collaborative data is {:.2E}".format(
data_sum, data_sum / data_nnz, collaborative_sum /
collaborative_nnz))
if self.evaluator_object is not None and self.show_max_performance:
self.computeMaxTheoreticalPerformance()