-
Notifications
You must be signed in to change notification settings - Fork 0
/
extra.py
195 lines (163 loc) · 8.13 KB
/
extra.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
import numpy as np
from scipy.stats import pearsonr
import matplotlib.pyplot as plt
import math
from sklearn.metrics import mean_squared_error, mean_absolute_error
MAX_GENERATIONS = 500
NUM_USERS = 943
NUM_MOVIES = 1682
rng = np.random.default_rng()
ratings_base = np.genfromtxt('ml-100k/ua.base', dtype='int32')
ratings_test = np.genfromtxt('ml-100k/ua.test', dtype='int32')
matrix_base = np.zeros((NUM_USERS, NUM_MOVIES), dtype='int32')
matrix_test = np.zeros((NUM_USERS, NUM_MOVIES), dtype='int32')
for r in ratings_base:
matrix_base[r[0]-1, r[1]-1] = r[2]
for r in ratings_test:
matrix_test[r[0] - 1, r[1] - 1] = r[2]
matrix_filled = matrix_base.copy()
unrated = np.where(matrix_filled == 0)
matrix_filled[unrated] = np.random.randint(low=1, high=6, size=len(unrated[1]))
rated_pos = np.where(matrix_base != 0)
user_ratings = []
neighborhoods = []
neighborhoods_ratings = []
for user in range(NUM_USERS):
ratings = np.where(matrix_base[user] != 0)
# Pair wise similarity with every user
similarity = np.array([pearsonr(matrix_filled[user][ratings], us[ratings])[0] for us in matrix_filled])
user_rated = matrix_filled[user][ratings]
similarity = np.empty(NUM_USERS)
for i, us in enumerate(matrix_filled):
us_rated = us[ratings]
pr = pearsonr(user_rated, us_rated)[0]
if math.isnan(pr):
pr = 0
similarity[i] = pr
max_users = similarity.argsort()[-11:-1][::-1] # Top 10 most similar users most similar one is user itself
neighborhood = matrix_base[max_users] # Neighborhood of user
# Actually rated movies of neighborhood
neighborhood_rated = [np.where(neigh != 0) for neigh in neighborhood]
neighborhoods.append(neighborhood)
neighborhoods_ratings.append(neighborhood_rated)
def fitness(x):
sum_pr = 0
for usr in range(NUM_USERS):
for i, neigh in enumerate(neighborhoods[usr]):
x_red = x[usr][neighborhoods_ratings[usr][i]]
neigh_red = neigh[neighborhoods_ratings[usr][i]]
pr = pearsonr(x_red, neigh_red)[0]
if math.isnan(pr):
pr = 0
sum_pr += pr
return sum_pr
# return sum([sum([
# pearsonr(x[usr][neighborhoods_ratings[usr][i]], neigh[neighborhoods_ratings[usr][i]])[0]
# for i, neigh in enumerate(neighborhoods[usr])]) for usr in range(NUM_USERS)])
def selection(pop_fitness):
num_pairs = int(POPULATION_SIZE * PROBABILITY_OF_CROSSOVER / 2) # Amount of parent pairs that should be selected
parents = np.empty((num_pairs, 2), dtype='int32')
for idx in range(num_pairs):
sample = rng.choice(POPULATION_SIZE, TOURNAMENT_SIZE, replace=False)
max_pair = pop_fitness[sample].argsort()[-2:][::-1]
parents[idx] = sample[max_pair]
return parents
def crossover(parent_pairs):
offspring = []
for parent in parent_pairs:
positions = np.random.randint(low=0, high=2, size=(NUM_USERS, NUM_MOVIES))
offspring.append(np.where(positions == 0, parent[0], parent[1]))
offspring.append(np.where(positions == 0, parent[1], parent[0]))
return offspring
def mutation(chromosomes):
num_mutations = int(PROBABILITY_OF_MUTATION * NUM_MOVIES * NUM_USERS)
for chrom in chromosomes:
positions = rng.choice(NUM_MOVIES * NUM_USERS, num_mutations, replace=False)
cols = (positions // NUM_USERS).reshape((1, num_mutations))
rows = (positions % NUM_USERS).reshape((1, num_mutations))
positions = np.concatenate((rows, cols), axis=0)
values = np.random.randint(1, 6, num_mutations)
chrom[tuple(positions)] = values
chrom[rated_pos] = matrix_base[rated_pos]
combs = [(0.6, 0.00, 20, 5), (0.6, 0.001, 20, 5), (0.6, 0.1, 20, 5), (0.9, 0.01, 20, 5), (0.1, 0.01, 20, 5),
(0.6, 0.00, 200, 20), (0.6, 0.01, 200, 20), (0.1, 0.01, 200, 20), (0.9, 0.01, 200, 20)]
print('RUN | Generations | Fitness | RMSE | MAE')
for ind, comb in enumerate(combs):
PROBABILITY_OF_CROSSOVER, PROBABILITY_OF_MUTATION, POPULATION_SIZE, TOURNAMENT_SIZE = comb
best = np.empty((MAX_GENERATIONS, 10))
rmse = np.empty((MAX_GENERATIONS, 10))
mae = np.empty((MAX_GENERATIONS, 10))
optimals = []
generations = []
for i in range(10):
# Create initial population
initial_population = np.random.randint(low=1, high=6, size=(POPULATION_SIZE, NUM_USERS, NUM_MOVIES))
for chromosome in initial_population:
chromosome[rated_pos] = matrix_base[rated_pos] # Inserts users actual ratings
next_pop = initial_population
next_pop_fitness = np.array([fitness(x) for x in next_pop])
next_sorted_keys = next_pop_fitness.argsort()
generation = 0
last_leader_change = 0
earlier_performance = 0
finished = False
while not finished:
current_pop = next_pop
current_pop_fitness = next_pop_fitness
current_sorted_keys = next_sorted_keys
parents = selection(current_pop_fitness)
children = crossover(current_pop[parents])
mutation(children)
next_pop = current_pop.copy()
next_pop[current_sorted_keys[:len(children)]] = np.array(children) # Replace worst solutions with new ones
children_fitness = np.array([fitness(x) for x in children])
next_pop_fitness = np.empty_like(current_pop_fitness)
next_pop_fitness[current_sorted_keys[len(children):]] = current_pop_fitness[current_sorted_keys[len(children):]]
next_pop_fitness[current_sorted_keys[:len(children)]] = children_fitness
next_sorted_keys = next_pop_fitness.argsort()
improvement = np.sum(next_pop_fitness) / np.sum(current_pop_fitness)
best_sol = next_pop[next_sorted_keys[-1]]
best_fitness = next_pop_fitness[next_sorted_keys[-1]]
best[generation][i] = best_fitness / NUM_USERS
rmse[generation][i] = np.sqrt(mean_squared_error(best_sol[rated_pos], matrix_test[rated_pos]))
mae[generation][i] = mean_absolute_error(best_sol[rated_pos], matrix_test[rated_pos])
finished = generation > MAX_GENERATIONS or generation - last_leader_change > 50
if not finished and generation >= 100:
finished = best[generation, i] / best[generation - 100, i] < 1.001
generation += 1
best[generation:, i] = max(next_pop_fitness)
rmse[generation:, i] = rmse[generation-1, i]
mae[generation:, i] = mae[generation-1, i]
optimals.append(max(next_pop_fitness))
generations.append(generation)
best_avg = np.average(best[:max(generations)], axis=1)
rmse_avg = np.average(rmse[:max(generations)], axis=1)
mae_avg = np.average(mae[:max(generations)], axis=1)
# print('RUN', ind)
# print('Average optimal', np.average(np.array(optimals)))
# print('Average generations', np.average(np.array(generations)))
# print('Average RMSE', rmse_avg[-1])
# print('Average MAE', mae_avg[-1])
print(ind, '&', np.average(np.array(generations)), '&', np.average(np.array(optimals)), '&', rmse_avg[-1], '&', mae_avg[-1], '\\\\')
# print('----------------------------')
plt.plot(np.arange(max(generations)), best_avg)
plt.title('Population size: ' + str(POPULATION_SIZE) + ' Crossover: ' + str(PROBABILITY_OF_CROSSOVER) + '% Mutation: '
+ str(PROBABILITY_OF_MUTATION) + '%')
plt.xlabel('Number of generations')
plt.ylabel('Average fitness of best value')
plt.grid(b=True, axis='y')
plt.show()
plt.plot(np.arange(max(generations)), rmse_avg)
plt.title('Population size: ' + str(POPULATION_SIZE) + ' Crossover: ' + str(PROBABILITY_OF_CROSSOVER) + '% Mutation: '
+ str(PROBABILITY_OF_MUTATION) + '%')
plt.xlabel('RMSE')
plt.ylabel('Average RMSE of best value')
plt.grid(b=True, axis='y')
plt.show()
plt.plot(np.arange(max(generations)), mae_avg)
plt.title('Population size: ' + str(POPULATION_SIZE) + ' Crossover: ' + str(PROBABILITY_OF_CROSSOVER) + '% Mutation: '
+ str(PROBABILITY_OF_MUTATION) + '%')
plt.xlabel('MAE')
plt.ylabel('Average MAE of best value')
plt.grid(b=True, axis='y')
plt.show()