def pcc_item_rating_pred(path, rating, method, k):
start = time.time()
name = 'pcc_item'
data = extract_data(path)
mtx = get_matrix(3).toarray()
item_mtx = []
result = []
zero = np.where(~mtx.any(axis=0))[0] # get zero
mtx[:, [zero]] = 0.00001 # prevent zero-devide
#normalize
pcc = (mtx.T - np.sum(mtx, axis=1)) / len(mtx)
pcc /= np.linalg.norm(mtx, axis=1).T
mtx = pcc.T
if method == 'dot':
item_mtx = dot_sim(mtx, name)
elif method == 'cos':
inputs = (mtx.T * np.linalg.norm(mtx, axis=1)).T
item_mtx = cos_sim(inputs, name)
#KNN
for i in data:
score = 0
item_id = i[0] #get item_id
user_id = i[1] #get user_id
item = item_mtx[item_id] #row
knn = np.argsort(item, kind='heapsort')[::-1][0:k + 1]
if item_id in knn: # delte query
idx = np.where(knn == item_id)
knn = np.delete(knn, idx)
else:
knn = np.delete(knn, len(knn) - 1)
#get score
if rating == 'mean':
score = np.sum(np.take(mtx[:, user_id],
knn.tolist())) / float(k) + 3
elif rating == 'weighted':
knn_sim = item[knn]
if np.sum(knn_sim) != 0: #prevent zero-devide
weight = knn_sim / np.sum(knn_sim)
score = np.sum(
np.multiply(np.take(mtx[:, user_id], knn.tolist()),
weight)) + 3
else:
score = np.sum(mtx[:, user_id]) / np.size(
np.nonzero(mtx[:, user_id])) + 3
result.append(score)
write(result, name, rating, method, k)
print('item_rating_pred {} {} {} time : {}'.format(method, rating, k,
time.time() - start))
gold = golden()
print("RMSE :", np.sqrt(np.mean(np.square(result - gold))))
def user_rating_pred(path, rating,method,k):
start = time.time()
name='user'
data = extract_data(path)
mtx = get_matrix(3).toarray()
user_mtx = []
result = []
zero = np.where(~mtx.any(axis=0))[0] #get zero
mtx[:, [zero]] = 0.00001 # prevent zero-devide
if method =='dot':
user_mtx = dot_sim(mtx,name)
elif method=='cos':
inputs=np.linalg.norm(mtx,axis=0)*mtx #normalize before cos_sim
user_mtx = cos_sim(inputs,name)#honestly cos_sim is cosine similariy but input is normalized so same with cos_similarity
for i in data:
score = 0
mv_id = i[0] #get item_id
user_id = i[1] #get user_id
user = user_mtx[user_id] #get user
knn = np.argsort(user,kind='heapsort')[::-1][0: k+1]
if user_id in knn:# delte query
i = np.where(knn == user_id)
knn = np.delete(knn, i)
else:
knn = np.delete(knn, len(knn) - 1)
#get score
if rating == 'mean':
score = (np.sum(np.take(mtx[mv_id, :], knn.tolist())) / float(k))+3
elif rating=='weighted':
knn_sim = user[knn]
if np.sum(knn_sim) != 0:
weight = knn_sim / np.sum(knn_sim) #prevent zero-devide
score = np.sum(np.multiply(np.take(mtx[mv_id, :], knn.tolist()), weight))+3
else:
score = np.sum(mtx[mv_id, :]) / np.size(np.nonzero(mtx[mv_id, :]))+3
result.append(score)
#print('start _writting')
write(result,name,rating,method,k)
print('user_rating_pred {} {} {} time : {}'.format(method,rating, k,time.time() - start))
gold=golden()
print("RMSE :",np.sqrt(np.mean(np.square(result-gold))))
#!/usr/bin/env python
# coding: utf-8
# In[2]:
from myutils import get_matrix, extract_data
import numpy as np
import torch
from copy import deepcopy
import time
# In[3]:
data = extract_data('data/dev.csv')
# In[4]:
import csv
f = open('data/dev.golden', 'r', encoding='utf-8')
reader = csv.reader(f)
golden = []
for i in reader:
golden += i
golden = np.array(golden, dtype=float)
# In[5]:
def get_score(U, V, data):
u = U.numpy().take(data.take(1, axis=1), axis=0)