duxy-me/convncf

Experimental codes for paper "Outer Product-based Neural Collaborative Filtering".

Outer Product-based Neural Collaborative Filtering

Convolutional Neural Collaborative Filtering performs well based on outer product of user and item embeddings. This is our official implementation for the paper:

Xiangnan He, Xiaoyu Du, Xiang Wang, Feng Tian, Jinhui Tang, Tat-Seng Chua, Outer Product-based Neural Collaborative Filtering, In Proceedings of IJCAI'18.

If you use the codes, please cite our paper . Thanks!

Requirements

• Tensorflow 1.7
• numpy, scipy

Quick Start

0. decompress the data files.

   cd Data
   gunzip *
   

1. Pretrain the embeddings using MF_BPR with

   python MF_BPR.py
   

2. Train ConvNCF with pretrained embeddings

   python ConvNCF.py --pretrain=1
   

Dataset

We provide the compressed dataset Yelp(yelp) in Data/

train.rating:

Train file.

Each Line is a training instance:

userID\t itemID\t rating\t timestamp (if have)

test.rating:

Test file (positive instances). Each Line is a testing instance:

userID\t itemID\t rating\t timestamp (if have)

test.negative

Test file (negative instances). Each line corresponds to the line of test.rating, containing 999 negative samples. Each line is in the format:

(userID,itemID)\t negativeItemID1\t negativeItemID2 ...

Files

• Data. Training and testing data.
  • yelp.train.rating. Rating of training data.
  • yelp.test.rating. Rating of testing data.
  • yelp.test.negative. 1000 testing samples for each user. (0,32) means this row is for user 0 and the positive test item is 32.
• Dataset.py. Module preprocessing data.
• saver.py. Module saving parameters.
• MF_BPR.py. MF model with BPR loss.
• ConvNCF.py. Our model.

convncf's Issues

The code cannot be executed

Hey! Thanks for sharing the code.

The code cannot be executed in my environment. And here are my configuration:
Windows 10, Anaconda, Python 3.5+Tensorflow 1.8

I have changed the ' print "" -> print() ' since the code seems to be written in python 2.x. After that, the code still cannot be executed, the errors are as follows:

1.NULL appears in the fifth line of saver.py. What does this mean?
2. When I run python MF_BPR.py, there is a bug.

Traceback (most recent call last):
  File "D:\Anaconda\lib\multiprocessing\pool.py", line 119, in worker
    result = (True, func(*args, **kwds))
  File "D:\Anaconda\lib\multiprocessing\pool.py", line 44, in mapstar
    return list(map(*args))
  File "E:\BaiduNetdiskDownload\ConvNCF-master\ConvNCF-master\MF_BPR.py", line 355, in _evaluate_input
    item_input = dataset.testNegatives[user] # read negative samples from files
NameError: name 'dataset' is not defined

3. Training ConvNCF with python ConvNCF.py --pretrain=0 also failed, the Traceback is

Traceback (most recent call last):
  File "ConvNCF.py", line 580, in <module>
    training(model, dataset, args)
  File "ConvNCF.py", line 319, in training
    eval_feed_dicts = init_eval_model(model, dataset)
  File "ConvNCF.py", line 444, in init_eval_model
    feed_dicts = pool.map(_evaluate_input, range(_dataset.num_users))
  File "D:\Anaconda\envs\tensorflow\lib\multiprocessing\pool.py", line 260, in map
    return self._map_async(func, iterable, mapstar, chunksize).get()
  File "D:\Anaconda\envs\tensorflow\lib\multiprocessing\pool.py", line 608, in get
    raise self._value
AttributeError: 'NoneType' object has no attribute 'testNegatives'

How can I fix it? Looking forward to your suggestion.

i have a question that what is the 'dimension correlations'?

Dataset

Please upload the script to process the dataset.

ConvNCF.py still load the evaluate model...

I study recently your project at https://github.com/duxy-me/ConvNCF.
Our team think your idea is awesome.

There are something to trouble you...
Hope you could help me.
I found some bug in your code when I run ConvNCF.py
Console is still show the screeshot at attachment
I hope sincerly you can relp me.

Best regards.
Kevin Xu

运行 python MF_BPR.py

Epoch 0 [2.2s + 10.6s]: HR = 0.1706, NDCG = 0.0905 AUC = 0.7705 train_AUC = 0.0000 [46.4s] ACC = 0.5003 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 50 [2.4s + 10.2s]: HR = 0.2545, NDCG = 0.1337 AUC = 0.9258 train_AUC = 0.0000 [32.8s] ACC = 0.9674 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 100 [2.4s + 10.1s]: HR = 0.2651, NDCG = 0.1379 AUC = 0.9295 train_AUC = 0.0000 [41.5s] ACC = 0.9730 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 150 [2.4s + 9.8s]: HR = 0.2695, NDCG = 0.1390 AUC = 0.9305 train_AUC = 0.0000 [41.5s] ACC = 0.9759 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 200 [2.4s + 9.9s]: HR = 0.2795, NDCG = 0.1474 AUC = 0.9318 train_AUC = 0.0000 [42.0s] ACC = 0.9777 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 250 [1.6s + 3.5s]: HR = 0.2801, NDCG = 0.1464 AUC = 0.9323 train_AUC = 0.0000 [18.6s] ACC = 0.9792 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 300 [1.7s + 3.5s]: HR = 0.2822, NDCG = 0.1485 AUC = 0.9330 train_AUC = 0.0000 [17.8s] ACC = 0.9798 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 350 [1.6s + 3.5s]: HR = 0.2875, NDCG = 0.1516 AUC = 0.9334 train_AUC = 0.0000 [18.3s] ACC = 0.9806 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 400 [1.7s + 3.5s]: HR = 0.2840, NDCG = 0.1491 AUC = 0.9336 train_AUC = 0.0000 [18.6s] ACC = 0.9807 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 450 [1.6s + 3.5s]: HR = 0.2841, NDCG = 0.1490 AUC = 0.9335 train_AUC = 0.0000 [17.7s] ACC = 0.9815 train_loss = 0.0000 ACC = 0.0000 [0.0s]
Epoch 499 [1.6s + 3.5s]: HR = 0.2860, NDCG = 0.1490 AUC = 0.9337 train_AUC = 0.0000 [18.4s] ACC = 0.9820 train_loss = 0.0000 ACC = 0.0000 [0.0s]
best:Epoch 350 [1.6s + 3.5s]: HR = 0.2875, NDCG = 0.1516 AUC = 0.9334 train_AUC = 0.0000 [18.3s] ACC = 0.9806 train_loss = 0.0000 ACC = 0.0000 [0.0s]

An Error "NameError: name 'dataset' is not defined" Occurred.

我的环境是win10+anconda3+tensorflow1.8，但是在跑代码时出现了这个错误，请问您知道该怎么解决吗？谢谢

An Error "NameError: name 'dataset' is not defined" Occurred.

Excuse me,when I attempt to run this paper's implements, some troubles caught me. Followed the "README.md", as I run "python MF_BPR.py", an Error occurs.

"NameError: name 'dataset' is not defined"

Train the ConvNCF

为什么interaction map可以解释为图像呢？

图像的局部是有相似/相同的语义的，所以用CNN来提取特征是有意义的。但interaction map只是两个embedding的外积，对interaction map来说，有这样的局部意义吗？请问你是怎么思考这一点的呢？谢谢！

代码里的dns和gtItem 表示的什么意思

conv-ncf

问下 为什么 optimizer 里设两个梯度？

the running time

hi,i hava a question that how much time do you spend running the ConvNCF.py,and your CPU and GPU?

The evaluate function is Pretty Confusing

ConvNCF/ConvNCF.py

Line 494 in 14674bc

def _eval_by_user(user):

I don't know why there are two return statement in the function, and the first function return the result of K=5,10,20 ?

I can't help wanting to diss the code~ maybe you can check more carefully before publish.