m.maheri
/
MetaTL
Watch 1
Star 0
Fork 0
Code Issues 0 Pull Requests 0 Releases 0 Wiki Activity
Browse Source

first commit

master
mohamad maheri 4 years ago
commit
9e1c651cfe
8 changed files with 402052 additions and 0 deletions
Split View Show Diff Stats
  1. 27
    0
      README.md
  2. 41357
    0
      data/electronics/electronics_test_new_user.csv
  3. 360069
    0
      data/electronics/electronics_train.csv
  4. 63
    0
      main.py
  5. 130
    0
      models.py
  6. 124
    0
      sampler.py
  7. 140
    0
      trainer.py
  8. 142
    0
      utils.py

+ 27
- 0
README.md View File

@@ -0,0 +1,27 @@
Sequential Recommendation for Cold-start Users with Meta Transitional Learning(SIGIR2021)
============

## CuRe

Code of paper "[Sequential Recommendation for Cold-start Users with Meta Transitional Learning](http://people.tamu.edu/~jwang713/pubs/MetaTL-sigir2021)".


## Requirements
python==3.6.8

## Usage
```python main.py --K 3```

## Cite

Please cite our paper if you use this code in your own work:

```
@inproceedings{wang2021sequential,
title={Sequential Recommendation for Cold-start Users with Meta Transitional Learning},
author={Wang, Jianling and Ding, Kaize and Caverlee, James},
booktitle={Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval},
pages={1783--1787},
year={2021}
}
```

+ 41357
- 0
data/electronics/electronics_test_new_user.csv
File diff suppressed because it is too large
View File


+ 360069
- 0
data/electronics/electronics_train.csv
File diff suppressed because it is too large
View File


+ 63
- 0
main.py View File

@@ -0,0 +1,63 @@
from trainer import *
from utils import *
from sampler import *
import json

import argparse


def get_params():
args = argparse.ArgumentParser()
args.add_argument("-data", "--dataset", default="electronics", type=str)
args.add_argument("-seed", "--seed", default=None, type=int)
args.add_argument("-K", "--K", default=3, type=int) #NUMBER OF SHOT

args.add_argument("-dim", "--embed_dim", default=100, type=int)
args.add_argument("-bs", "--batch_size", default=1024, type=int)
args.add_argument("-lr", "--learning_rate", default=0.001, type=float)

args.add_argument("-epo", "--epoch", default=100000, type=int)
args.add_argument("-prt_epo", "--print_epoch", default=100, type=int)
args.add_argument("-eval_epo", "--eval_epoch", default=1000, type=int)

args.add_argument("-b", "--beta", default=5, type=float)
args.add_argument("-m", "--margin", default=1, type=float)
args.add_argument("-p", "--dropout_p", default=0.5, type=float)

args.add_argument("-gpu", "--device", default=0, type=int)


args = args.parse_args()
params = {}
for k, v in vars(args).items():
params[k] = v

params['device'] = torch.device('cuda:'+str(args.device))

return params, args

if __name__ == '__main__':
params, args = get_params()

if params['seed'] is not None:
SEED = params['seed']
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
np.random.seed(SEED)
random.seed(SEED)

user_train, usernum_train, itemnum, user_input_test, user_test, user_input_valid, user_valid = data_load(args.dataset, args.K)

sampler = WarpSampler(user_train, usernum_train, itemnum, batch_size=args.batch_size, maxlen=args.K, n_workers=3)

sampler_test = DataLoader(user_input_test, user_test, itemnum, params)

sampler_valid = DataLoader(user_input_valid, user_valid, itemnum, params)

trainer = Trainer([sampler, sampler_valid, sampler_test], itemnum, params)

trainer.train()

sampler.close()

+ 130
- 0
models.py View File

@@ -0,0 +1,130 @@
from collections import OrderedDict
import torch
import torch.nn as nn
from torch.nn import functional as F

class Embedding(nn.Module):
def __init__(self, num_ent, parameter):
super(Embedding, self).__init__()
self.device = parameter['device']
self.es = parameter['embed_dim']
self.embedding = nn.Embedding(num_ent + 1, self.es)
nn.init.xavier_uniform_(self.embedding.weight)


def forward(self, triples):
idx = [[[t[0], t[2]] for t in batch] for batch in triples]
idx = torch.LongTensor(idx).to(self.device)
return self.embedding(idx)

class MetaLearner(nn.Module):
def __init__(self, K, embed_size=100, num_hidden1=500, num_hidden2=200, out_size=100, dropout_p=0.5):
super(MetaLearner, self).__init__()
self.embed_size = embed_size
self.K = K
self.out_size = out_size
self.rel_fc1 = nn.Sequential(OrderedDict([
('fc', nn.Linear(2*embed_size, num_hidden1)),
('bn', nn.BatchNorm1d(K)),
('relu', nn.LeakyReLU()),
('drop', nn.Dropout(p=dropout_p)),
]))
self.rel_fc2 = nn.Sequential(OrderedDict([
('fc', nn.Linear(num_hidden1, num_hidden2)),
('bn', nn.BatchNorm1d(K)),
('relu', nn.LeakyReLU()),
('drop', nn.Dropout(p=dropout_p)),
]))
self.rel_fc3 = nn.Sequential(OrderedDict([
('fc', nn.Linear(num_hidden2, out_size)),
('bn', nn.BatchNorm1d(K)),
]))
nn.init.xavier_normal_(self.rel_fc1.fc.weight)
nn.init.xavier_normal_(self.rel_fc2.fc.weight)
nn.init.xavier_normal_(self.rel_fc3.fc.weight)

def forward(self, inputs):
size = inputs.shape
x = inputs.contiguous().view(size[0], size[1], -1)
x = self.rel_fc1(x)
x = self.rel_fc2(x)
x = self.rel_fc3(x)
x = torch.mean(x, 1)

return x.view(size[0], 1, 1, self.out_size)


class EmbeddingLearner(nn.Module):
def __init__(self):
super(EmbeddingLearner, self).__init__()

def forward(self, h, t, r, pos_num):
score = -torch.norm(h + r - t, 2, -1).squeeze(2)
p_score = score[:, :pos_num]
n_score = score[:, pos_num:]
return p_score, n_score


class MetaTL(nn.Module):
def __init__(self, itemnum, parameter):
super(MetaTL, self).__init__()
self.device = parameter['device']
self.beta = parameter['beta']
self.dropout_p = parameter['dropout_p']
self.embed_dim = parameter['embed_dim']
self.margin = parameter['margin']
self.embedding = Embedding(itemnum, parameter)

self.relation_learner = MetaLearner(parameter['K'] - 1, embed_size=100, num_hidden1=500,
num_hidden2=200, out_size=100, dropout_p=self.dropout_p)

self.embedding_learner = EmbeddingLearner()
self.loss_func = nn.MarginRankingLoss(self.margin)
self.rel_q_sharing = dict()

def split_concat(self, positive, negative):
pos_neg_e1 = torch.cat([positive[:, :, 0, :],
negative[:, :, 0, :]], 1).unsqueeze(2)
pos_neg_e2 = torch.cat([positive[:, :, 1, :],
negative[:, :, 1, :]], 1).unsqueeze(2)
return pos_neg_e1, pos_neg_e2

def forward(self, task, iseval=False, curr_rel=''):
# transfer task string into embedding
support, support_negative, query, negative = [self.embedding(t) for t in task]

K = support.shape[1] # num of K
num_sn = support_negative.shape[1] # num of support negative
num_q = query.shape[1] # num of query
num_n = negative.shape[1] # num of query negative

rel = self.relation_learner(support)
rel.retain_grad()

rel_s = rel.expand(-1, K+num_sn, -1, -1)

if iseval and curr_rel != '' and curr_rel in self.rel_q_sharing.keys():
rel_q = self.rel_q_sharing[curr_rel]
else:
sup_neg_e1, sup_neg_e2 = self.split_concat(support, support_negative)

p_score, n_score = self.embedding_learner(sup_neg_e1, sup_neg_e2, rel_s, K)

y = torch.Tensor([1]).to(self.device)
self.zero_grad()
loss = self.loss_func(p_score, n_score, y)
loss.backward(retain_graph=True)

grad_meta = rel.grad
rel_q = rel - self.beta*grad_meta


self.rel_q_sharing[curr_rel] = rel_q

rel_q = rel_q.expand(-1, num_q + num_n, -1, -1)

que_neg_e1, que_neg_e2 = self.split_concat(query, negative)
p_score, n_score = self.embedding_learner(que_neg_e1, que_neg_e2, rel_q, num_q)

return p_score, n_score

+ 124
- 0
sampler.py View File

@@ -0,0 +1,124 @@
import sys
import copy
import torch
import random
import numpy as np
from collections import defaultdict, Counter
from multiprocessing import Process, Queue


def random_neq(l, r, s):
t = np.random.randint(l, r)
while t in s:
t = np.random.randint(l, r)
return t

def sample_function_mixed(user_train, usernum, itemnum, batch_size, maxlen, result_queue, SEED):
def sample():

if random.random()<0.5:
user = np.random.randint(1, usernum + 1)
while len(user_train[user]) <= 1: user = np.random.randint(1, usernum + 1)

seq = np.zeros([maxlen], dtype=np.int32)
pos = np.zeros([maxlen], dtype=np.int32)
neg = np.zeros([maxlen], dtype=np.int32)

if len(user_train[user]) < maxlen:
nxt_idx = len(user_train[user]) - 1
else:
nxt_idx = np.random.randint(maxlen,len(user_train[user]))

nxt = user_train[user][nxt_idx]
idx = maxlen - 1

ts = set(user_train[user])
for i in reversed(user_train[user][min(0, nxt_idx - 1 - maxlen) : nxt_idx - 1]):
seq[idx] = i
pos[idx] = nxt
if nxt != 0: neg[idx] = random_neq(1, itemnum + 1, ts)
nxt = i
idx -= 1
if idx == -1: break

curr_rel = user
support_triples, support_negative_triples, query_triples, negative_triples = [], [], [], []
for idx in range(maxlen-1):
support_triples.append([seq[idx],curr_rel,pos[idx]])
support_negative_triples.append([seq[idx],curr_rel,neg[idx]])
query_triples.append([seq[-1],curr_rel,pos[-1]])
negative_triples.append([seq[-1],curr_rel,neg[-1]])

return support_triples, support_negative_triples, query_triples, negative_triples, curr_rel

else:
user = np.random.randint(1, usernum + 1)
while len(user_train[user]) <= 1: user = np.random.randint(1, usernum + 1)

seq = np.zeros([maxlen], dtype=np.int32)
pos = np.zeros([maxlen], dtype=np.int32)
neg = np.zeros([maxlen], dtype=np.int32)

list_idx = random.sample([i for i in range(len(user_train[user]))], maxlen + 1)
list_item = [user_train[user][i] for i in sorted(list_idx)]

nxt = list_item[-1]
idx = maxlen - 1

ts = set(user_train[user])
for i in reversed(list_item[:-1]):
seq[idx] = i
pos[idx] = nxt
if nxt != 0: neg[idx] = random_neq(1, itemnum + 1, ts)
nxt = i
idx -= 1
if idx == -1: break

curr_rel = user
support_triples, support_negative_triples, query_triples, negative_triples = [], [], [], []
for idx in range(maxlen-1):
support_triples.append([seq[idx],curr_rel,pos[idx]])
support_negative_triples.append([seq[idx],curr_rel,neg[idx]])
query_triples.append([seq[-1],curr_rel,pos[-1]])
negative_triples.append([seq[-1],curr_rel,neg[-1]])

return support_triples, support_negative_triples, query_triples, negative_triples, curr_rel

np.random.seed(SEED)
while True:
one_batch = []
for i in range(batch_size):
one_batch.append(sample())

support, support_negative, query, negative, curr_rel = zip(*one_batch)

result_queue.put(([support, support_negative, query, negative], curr_rel))

class WarpSampler(object):
def __init__(self, User, usernum, itemnum, batch_size=64, maxlen=10, n_workers=1):
self.result_queue = Queue(maxsize=n_workers * 10)
self.processors = []
for i in range(n_workers):
self.processors.append(
Process(target=sample_function_mixed, args=(User,
usernum,
itemnum,
batch_size,
maxlen,
self.result_queue,
np.random.randint(2e9)
)))
self.processors[-1].daemon = True
self.processors[-1].start()

def next_batch(self):
return self.result_queue.get()

def close(self):
for p in self.processors:
p.terminate()
p.join()



+ 140
- 0
trainer.py View File

@@ -0,0 +1,140 @@
from models import *
import os
import sys
import torch
import shutil
import logging
import numpy as np


class Trainer:
def __init__(self, data_loaders, itemnum, parameter):
self.parameter = parameter
# data loader
self.train_data_loader = data_loaders[0]
self.dev_data_loader = data_loaders[1]
self.test_data_loader = data_loaders[2]
# parameters
self.batch_size = parameter['batch_size']
self.learning_rate = parameter['learning_rate']
self.epoch = parameter['epoch']
self.print_epoch = parameter['print_epoch']
self.eval_epoch = parameter['eval_epoch']
self.device = parameter['device']

self.MetaTL = MetaTL(itemnum, parameter)
self.MetaTL.to(self.device)

self.optimizer = torch.optim.Adam(self.MetaTL.parameters(), self.learning_rate)

def rank_predict(self, data, x, ranks):
# query_idx is the idx of positive score
query_idx = x.shape[0] - 1
# sort all scores with descending, because more plausible triple has higher score
_, idx = torch.sort(x, descending=True)
rank = list(idx.cpu().numpy()).index(query_idx) + 1
ranks.append(rank)
# update data
if rank <= 10:
data['Hits@10'] += 1
data['NDCG@10'] += 1 / np.log2(rank + 1)
if rank <= 5:
data['Hits@5'] += 1
data['NDCG@5'] += 1 / np.log2(rank + 1)
if rank == 1:
data['Hits@1'] += 1
data['NDCG@1'] += 1 / np.log2(rank + 1)
data['MRR'] += 1.0 / rank

def do_one_step(self, task, iseval=False, curr_rel=''):
loss, p_score, n_score = 0, 0, 0
if not iseval:
self.optimizer.zero_grad()
p_score, n_score = self.MetaTL(task, iseval, curr_rel)
y = torch.Tensor([1]).to(self.device)
loss = self.MetaTL.loss_func(p_score, n_score, y)
loss.backward()
self.optimizer.step()
elif curr_rel != '':
p_score, n_score = self.MetaTL(task, iseval, curr_rel)
y = torch.Tensor([1]).to(self.device)
loss = self.MetaTL.loss_func(p_score, n_score, y)
return loss, p_score, n_score

def train(self):
# initialization
best_epoch = 0
best_value = 0
bad_counts = 0

# training by epoch
for e in range(self.epoch):
# sample one batch from data_loader
train_task, curr_rel = self.train_data_loader.next_batch()
loss, _, _ = self.do_one_step(train_task, iseval=False, curr_rel=curr_rel)
# print the loss on specific epoch
if e % self.print_epoch == 0:
loss_num = loss.item()
print("Epoch: {}\tLoss: {:.4f}".format(e, loss_num))
# do evaluation on specific epoch
if e % self.eval_epoch == 0 and e != 0:
print('Epoch {} Validating...'.format(e))
valid_data = self.eval(istest=False, epoch=e)

print('Epoch {} Testing...'.format(e))
test_data = self.eval(istest=True, epoch=e)
print('Finish')

def eval(self, istest=False, epoch=None):
self.MetaTL.eval()
self.MetaTL.rel_q_sharing = dict()

if istest:
data_loader = self.test_data_loader
else:
data_loader = self.dev_data_loader
data_loader.curr_tri_idx = 0

# initial return data of validation
data = {'MRR': 0, 'Hits@1': 0, 'Hits@5': 0, 'Hits@10': 0, 'NDCG@1': 0, 'NDCG@5': 0, 'NDCG@10': 0}
ranks = []

t = 0
temp = dict()
while True:
# sample all the eval tasks
eval_task, curr_rel = data_loader.next_one_on_eval()
# at the end of sample tasks, a symbol 'EOT' will return
if eval_task == 'EOT':
break
t += 1

_, p_score, n_score = self.do_one_step(eval_task, iseval=True, curr_rel=curr_rel)

x = torch.cat([n_score, p_score], 1).squeeze()

self.rank_predict(data, x, ranks)

# print current temp data dynamically
for k in data.keys():
temp[k] = data[k] / t
sys.stdout.write("{}\tMRR: {:.3f}\tNDCG@10: {:.3f}\tNDCG@5: {:.3f}\tNDCG@1: {:.3f}\tHits@10: {:.3f}\tHits@5: {:.3f}\tHits@1: {:.3f}\r".format(
t, temp['MRR'], temp['NDCG@10'], temp['NDCG@5'], temp['NDCG@1'], temp['Hits@10'], temp['Hits@5'], temp['Hits@1']))
sys.stdout.flush()

# print overall evaluation result and return it
for k in data.keys():
data[k] = round(data[k] / t, 3)

if istest:
print("TEST: \tMRR: {:.3f}\tNDCG@10: {:.3f}\tNDCG@5: {:.3f}\tNDCG@1: {:.3f}\tHits@10: {:.3f}\tHits@5: {:.3f}\tHits@1: {:.3f}\r".format(
temp['MRR'], temp['NDCG@10'], temp['NDCG@5'], temp['NDCG@1'], temp['Hits@10'], temp['Hits@5'], temp['Hits@1']))
else:
print("VALID: \tMRR: {:.3f}\tNDCG@10: {:.3f}\tNDCG@5: {:.3f}\tNDCG@1: {:.3f}\tHits@10: {:.3f}\tHits@5: {:.3f}\tHits@1: {:.3f}\r".format(
temp['MRR'], temp['NDCG@10'], temp['NDCG@5'], temp['NDCG@1'], temp['Hits@10'], temp['Hits@5'], temp['Hits@1']))

return data

+ 142
- 0
utils.py View File

@@ -0,0 +1,142 @@
import sys
import copy
import torch
import random
import numpy as np
from collections import defaultdict, Counter
from multiprocessing import Process, Queue
# sampler for batch generation
def random_neq(l, r, s):
t = np.random.randint(l, r)
while t in s:
t = np.random.randint(l, r)
return t

def trans_to_cuda(variable):
if torch.cuda.is_available():
return variable.cuda()
else:
return variable


def trans_to_cpu(variable):
if torch.cuda.is_available():
return variable.cpu()
else:
return variable

# train/val/test data generation
def data_load(fname, num_sample):
usernum = 0
itemnum = 0
user_train = defaultdict(list)

# assume user/item index starting from 1
f = open('data/%s/%s_train.csv' % (fname, fname), 'r')
for line in f:
u, i, t = line.rstrip().split('\t')
u = int(u)
i = int(i)
usernum = max(u, usernum)
itemnum = max(i, itemnum)
user_train[u].append(i)
f.close()

# read in new users for testing
user_input_test = {}
user_input_valid = {}
user_valid = {}
user_test = {}


User_test_new = defaultdict(list)
f = open('data/%s/%s_test_new_user.csv' % (fname, fname), 'r')
for line in f:
u, i, t = line.rstrip().split('\t')
u = int(u)
i = int(i)
User_test_new[u].append(i)
f.close()

for user in User_test_new:
if len(User_test_new[user]) > num_sample:
if random.random()<0.3:
user_input_valid[user] = User_test_new[user][:num_sample]
user_valid[user] = []
user_valid[user].append(User_test_new[user][num_sample])
else:
user_input_test[user] = User_test_new[user][:num_sample]
user_test[user] = []
user_test[user].append(User_test_new[user][num_sample])

return [user_train, usernum, itemnum, user_input_test, user_test, user_input_valid, user_valid]







class DataLoader(object):
def __init__(self, user_train, user_test, itemnum, parameter):
self.curr_rel_idx = 0
self.bs = parameter['batch_size']
self.maxlen = parameter['K']

self.valid_user = []
for u in user_train:
if len(user_train[u]) < self.maxlen or len(user_test[u]) < 1: continue
self.valid_user.append(u)
self.num_tris = len(self.valid_user)

self.train = user_train
self.test = user_test
self.itemnum = itemnum

def next_one_on_eval(self):
if self.curr_tri_idx == self.num_tris:
return "EOT", "EOT"

u = self.valid_user[self.curr_tri_idx]

self.curr_tri_idx += 1
seq = np.zeros([self.maxlen], dtype=np.int32)
pos = np.zeros([self.maxlen - 1], dtype=np.int32)
neg = np.zeros([self.maxlen - 1], dtype=np.int32)
idx = self.maxlen - 1

ts = set(self.train[u])
for i in reversed(self.train[u]):
seq[idx] = i
if idx > 0:
pos[idx - 1] = i
if i != 0: neg[idx - 1] = random_neq(1, self.itemnum + 1, ts)
idx -= 1
if idx == -1: break

curr_rel = u
support_triples, support_negative_triples, query_triples, negative_triples = [], [], [], []
for idx in range(self.maxlen-1):
support_triples.append([seq[idx],curr_rel,pos[idx]])
support_negative_triples.append([seq[idx],curr_rel,neg[idx]])

rated = ts
rated.add(0)
query_triples.append([seq[-1],curr_rel,self.test[u][0]])
for _ in range(100):
t = np.random.randint(1, self.itemnum + 1)
while t in rated: t = np.random.randint(1, self.itemnum + 1)
negative_triples.append([seq[-1],curr_rel,t])

support_triples = [support_triples]
support_negative_triples = [support_negative_triples]
query_triples = [query_triples]
negative_triples = [negative_triples]

return [support_triples, support_negative_triples, query_triples, negative_triples], curr_rel

Write Preview
Loading…
Cancel
Save