zyang1580/sml

This is an implementation for our SIGIR 2020 paper: How to Retrain Recommender System? A Sequential Meta-Learning Method.

sml's Issues

About Adressa Dataset

Thanks for all your hard work!
May I have a chance to know that do we have Adressa (news) dataset? We haven't seen it in the netdisk.

About the meta learner

It's really a fantastic work！！ I‘m new to meta learning，and I have spend lots of time to understand meta learning these days， but I didn't figure out which kind of meta learning this work belongs to？Feed-Forword Model？Black-Box？or which work before give this spike to you？look forward to you reply

About dataset

SML/data/dataset.py

Line 95 in 43f49d5

information = np.load(self.path+self.dataname+'/'+"information.npy")

Hi, Thanks for your contributions~ and i have some questions about dataset and algorithm in your papers.

1. dataset of yelp is in JSON format, and seems different from your data format. Can you provide your pre-process code or some sampe-examples?
2. In algorithm2, why the \hat(W_t) is updated twice with D_t ? which has been updated in algorithm1 with D_t.

Thanks.

About Pre-train model in save_model

May I ask about another issue? I wonder how you pre-trained the model, and I can't find any information about it in the paper.
What I've done is that I used the most effective full-retrain model (after tuning the hyperparameters) as the pretrained one (baseline_init), for training fine-tune method, the recall@20 will be much higher than that proposed in the paper, which was 0.86 or higher.
If I used gowalla dataset, it will reach to 0.95, which is very incredible.

About the user and item amount of Yelp

I found that in your paper, the user and item of Yelp is 59082, 122816, respectively. However, in your code, there are 122816 users and 59082 items, and in the .npy files, users are encoded in [0,122815], while items are encoded in [0,59081]. May I ask is the number in the paper reversed? Or the user and item has put reversely in your code?

Create neg_item from input dataset

SML/data/dataset2.py

Lines 172 to 201 in 2e1e3ee

	class trainDataset_withPreSample(Dataset):
	'''
	this is Dataset type for train transfer, the input dataset has sampled enough
	neg item. the each epoch, will select on cloumn neg_item as neg item
	'''
	def __init__(self,input_dataset):
	super(trainDataset_withPreSample, self).__init__()
	self.all_data = copy.deepcopy(input_dataset)
	self.have_read = 0
	self.neg_flag = np.arange(1, self.all_data.shape[1])
	np.random.shuffle(self.neg_flag)
	self.neg_all = input_dataset.shape[1]-2
	self.used_neg_count = 0
	self.data_len = input_dataset.shape[0]
	def __len__(self):
	return self.all_data.shape[0]
	def __getitem__(self, idx):
	user = self.all_data[idx,0]
	item = self.all_data[idx,1]
	neg_item = self.all_data[idx,self.neg_flag[self.used_neg_count]]
	self.have_read += 1
	if self.have_read >= self.data_len:
	self.have_read = 0
	self.used_neg_count += 1
	if self.used_neg_count >= self.neg_all:
	np.random.shuffle(self.neg_flag)
	self.used_neg_count = 0
	return user,item,neg_item

Hi, thank you for your contribution. I have a question regarding the way you generate neg_item from input dataset. As for what I understand from readme and your code comments, the raw input dataset is of the format (user_id, pos_item_id, neg_item_id_1, neg_item_id_2 ... neg_item_id_n), hence the neg_item_id starts from third column. However, the self.neg_flag defined in the above class starts from 1, which is the second column. Can you help explain your code here? Please feel free to correct me if I'm wrong.

Thank you!