-
Notifications
You must be signed in to change notification settings - Fork 0
/
simKNN.py
168 lines (148 loc) · 5.89 KB
/
simKNN.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
import numpy as np
from tqdm.notebook import tqdm
class SimKNN:
def __init__(self, k, rho=0.4, metric="cosine", weight="uniform", verbose=True, most=100, debug=False):
'''
k : int \\
rho : float; 0.3(default) only for idf \\
metric : string; "cosine"(default), "amplification", "idf", "position", etc or function \\
weight : string; "uniform"(default), "distance", etc or fucntion \\
verbose : boolean
'''
self.data = None
self.data_id = None
self.X = None
self.X_id = None
self.freq = None
# hyperparameters
self.k = k
self.rho = rho
# ---
self.metric = metric
self.weight = weight # TODO
self.verbose = verbose
self.most = most
self.debug = debug
self.__version__ = "1.0"
self._checker()
def fit(self, data):
'''
data : pandas.DataFrame (columns=['id', 'songs'])
'''
self.data = data.iloc[:, 1].apply(np.array).to_numpy()
self.data_id = data.iloc[:, 0].copy(); del data
if self.metric == "idf":
self.freq = np.zeros(707989, dtype=np.int64)
_data = tqdm(self.data) if self.verbose else self.data
for datum in _data:
self.freq[datum] += 1
def predict(self, X, generator=False, limit=None, start=None, end=None):
'''
parameters \\
X : pandas.DataFrame (columns=['id', 'songs']) \\
returns \\
list
'''
self.X = X.iloc[:, 1].apply(np.array).to_numpy()
self.X_id = X.iloc[:, 0].copy(); del X
pred = []
if type(limit) == int and limit > 0:
_range = tqdm(range(limit)) if self.verbose else range(limit)
elif end:
_range = tqdm(range(start, end)) if self.verbose else range(start, end)
else:
_range = tqdm(range(self.X.size)) if self.verbose else range(self.X.size)
for u in _range: # FIXME : double for loops -> numpy broadcasting?
S = np.array([self._sim(u, v) for v in range(self.data.size)]) # sim of row for u and v
if self.debug:
top = S.argsort()[-(self.k+1):-1]
else:
top = S.argsort()[-self.k:]
norm = S[top].sum()
tracks = np.unique(np.concatenate(self.data[top])) # all tracks in top k playlists
tracks = np.setdiff1d(tracks, self.X[u], assume_unique=True) # remove common songs
r_u = np.array([( track, np.sum([S[v] if track in self.data[v] else 0 for v in top]) / norm) for track in tracks])
# L r_u_hat / FIXME : double for loops
# TODO : add weight
del S, top, norm, tracks
r_u = r_u[r_u[:, 1].argsort()][::-1][:self.most]
# yield u, r_u[:, 0].astype(np.int64), r_u[:, 1]
r_u = (u, r_u[:, 0].astype(np.int64), r_u[:, 1])
# tuple (playlist order, predicted songs id, relevance)
pred.append(r_u)
return pred
def _sim(self, u, v):
'''
u : int; u is index of playlist in test.json \\
v : int; v is index of playlist in train.json
'''
if self.X[u].size == 0:
return 0
elif self.metric == "cosine":
u = self.X[u] # numpy array
v = self.data[v] # numpy array
return np.intersect1d(u, v).size / ((u.size ** 0.5) * (v.size ** 0.5))
elif self.metric == "idf":
u = self.X[u]
v = self.data[v]
freq = self.freq[np.intersect1d(u, v)]
freq = 1 / (((freq - 1) ** self.rho) + 1)
return freq.sum() / ((u.size ** 0.5) * (v.size ** 0.5))
# {{ add other similarity metrics here }}
elif hasattr(self.metric, '__call__'):
return self.metric(u, v)
# def score(self, data, split=0.3, limit=limit):
# '''
# data : pandas.DataFrame (columns=['id', 'songs'])
# '''
# size = int(data.shape[0] * (1 - split))
# X = data.iloc[size:, :] # TODO split label
# y = data.iloc[size:, :] # TODO
# data = data.iloc[size, :]
# self.fit(data)
# pred = self.predict(X, limit=limit)
# return self._ndcg(y, pred)
# def _ndcg(self, y, pred):
# dcg = 0.0
# for i, r in enumerate(pred):
# if r in y:
# dcg += 1.0 / np.log(i + 2)
# return dcg / self._idcgs[len(y)]
# def _idcg(self, l):
# return sum((1.0 / np.log(i + 2) for i in range(l)))
# def _remove_common_songs(self, u, r_u):
# u = self.X[u]
# _most = self.most + u.size
# r_u = r_u[r_u[:, 1].argsort()][::-1][:_most] # select top 100 + alpha
# tmp = r_u[:, 0].astype(np.int64)
# tmp = np.setdiff1d(tmp, u, assume_unique=True)
# return np.array([pair for pair in r_u if pair[0] in tmp])[:self.most]
# # pair = (track_id, relevance) TODO : for loop -> numpy
def _checker(self): # error checker, save your time!
if type(self.k) == type(1):
pass
else:
raise TypeError
metric_keys = ["cosine", "amplification", "idf", "position"]
if self.metric in metric_keys:
pass
elif hasattr(self.metric, '__call__'):
pass
else:
raise KeyError("invalid key : {}".format(self.metric))
# TODO: remove weight
weight_keys = ["uniform", "distance"]
if self.weight in weight_keys:
pass
elif hasattr(self.weight, '__call__'):
pass
else:
raise KeyError("invalid key : {}".format(self.weight))
if __name__=='__main__':
import pickle
with open("bin/Xs.p", 'rb') as f:
Xs = pickle.load(f)
simknn = SimKNN(k=100, metric="idf")
simknn.fit(Xs[0][['id', 'songs']])
pred = simknn.predict(Xs[1][['id', 'songs']], limit=1)
print(pred)