def calculate_similarity(user_feature, article_feature):
# NOTE: user_feature is matrix, not support bool operation
if user_feature is None:
return 0
rst = _cosine_similarity(user_feature, article_feature)
if not rst:
return 0
return int(rst[0, 0] * 1000)
def batch_calculate_similarity(user_feature, article_matrix):
# NOTE: user_feature is matrix, not support bool operation
if user_feature is None:
return xrange(article_matrix.shape[0])
sims_matrix = _cosine_similarity(user_feature, article_matrix)
return sims_matrix[0,].tolist()
def cosine_similarity(x, y):
return _cosine_similarity(x, y)[0, 0]
def sub_set_coll_scores(review_set, review_hist, users, restaurants):
count = 1
tot = review_set.shape[0]
rest_id = None
for rid, row in review_set.iterrows():
user_id = row['user_id']
curr_user = users.loc[user_id]
old_rest_id = rest_id
rest_id = row['business_id']
if old_rest_id != rest_id:
review_rest_new = review_set.loc[review_set.business_id == rest_id, rev_cols]
review_rest_old = review_hist.loc[review_hist.business_id == rest_id, rev_cols]
tmp_review_rest = _pd.concat([review_rest_new, review_rest_old])
tmp_review_rest = tmp_review_rest.groupby('user_id').apply(aggregate)
tmp_user_rest = users.loc[users.index.isin(tmp_review_rest.index)]
review_rest = tmp_review_rest.drop(user_id)
user_rest = tmp_user_rest.drop(user_id)
assert review_rest.shape[0] == user_rest.shape[0], "different shapes: " + str(review_rest.shape) + " vs " + str(user_rest.shape)
a_u = row['cuisine_av_hist']
a_u_bin = row['cuisine_av_hist_bin']
a_u_real = row['cuisine_av_hist_real']
if user_rest.empty:
res = 0
res_bin = 0
res_real = 0
else:
a_r = restaurants.loc[rest_id, 'average_stars']
a_u_r = review_rest['stars']
user_sim = _cosine_similarity(curr_user[cols_std].values.reshape(1, -1), user_rest[cols_std])
user_sim = _pd.Series(data=user_sim[0], index=user_rest.index)
user_sim.where(user_sim > 0.5, 0, inplace=True)
numerator = (user_sim * (a_u_r - a_r)).sum()
denominator = user_sim.sum()
res = numerator / denominator
a_r_bin = restaurants.loc[rest_id, 'average_stars_bin']
a_u_r_bin = review_rest['stars_bin'].fillna(a_r_bin)
user_sim = _cosine_similarity(curr_user[cols_bin].values.reshape(1, -1), user_rest[cols_bin])
user_sim = _pd.Series(data=user_sim[0], index=user_rest.index)
user_sim.where(user_sim > 0.5, 0, inplace=True)
numerator_bin = (user_sim * (a_u_r_bin - a_r_bin)).sum()
denominator_bin = user_sim.sum()
res_bin = numerator_bin / denominator_bin
a_r_real = restaurants.loc[rest_id, 'average_stars_real']
a_u_r_real = review_rest['stars_real']
user_sim = _cosine_similarity(curr_user[cols_real].values.reshape(1, -1), user_rest[cols_real])
user_sim = _pd.Series(data=user_sim[0], index=user_rest.index)
user_sim.where(user_sim > 0.5, 0, inplace=True)
numerator_real = (user_sim * (a_u_r_real - a_r_real)).sum()
denominator_real = user_sim.sum()
res_real = numerator_real / denominator_real
out_cols = ['coll_score', 'coll_score_bin', 'coll_score_real']
vals = [a_u + res, a_u_bin + res_bin, a_u_real + res_real]
review_set.loc[rid, out_cols] = vals
if count % 1000 == 0:
percent = (count / tot) * 100
print("process {4}\t- row {0}/{1}\t- {2:.3f}%\t- {3}"
.format(count, tot, percent, _time.asctime(), _os.getpid()))
count += 1
