newlei / fairgo Goto Github PK

您好！拜读了您的paper和部分code，感受颇深，我想尝试复现效果并在此基础上做一些改进。但是我从您的代码和数据集链接中并未找到完整的数据集处理文件（只看到一小部分处理lastfm的data.py，而且还很不完整，只是提取了user和item的属性部分），希望您可以上传一下处理数据集的py文件，感谢！！！

您好！在您的paper中L-th order egocentric 公式里面会涉及到评分矩阵A的元素，我的疑问是既然任务是要预测rating，为什么还可以在训练过程中以引入rating呢？这并不合理吧？

Hi, I am reading your paper and codes, it is very interesting work and I would like to take your work as a baseline. But I have some questions about the implementation.

(1) when updating the filters, it seems that there are two goals -- minimize the recommendation loss and maximize the classification loss. In your paper, a hyperparameter $\lambda$ is used to combine the two losses together to realize the goals. But in your codes, it seems that the two corresponding losses are optimized separately, and the classification is only updated one batch as following. Which one should I make use of?

for user_batch, rating_batch, item_batch in train_loader: 
        user_batch = user_batch.cuda()
        rating_batch = rating_batch.cuda()
        item_batch = item_batch.cuda()
        d_g_l_get = model(copy.deepcopy(pos_adj),user_batch,rating_batch,item_batch)
        # d_g_l_get = model(copy.deepcopy(pos_adj),copy.deepcopy(pos_adj),user,item_i, item_j) 
        _,f_l,d_l = d_g_l_get 
        loss_current[0].append(f_l.item()) 
        # loss_current[1].append(d_l.item())  
        f_optimizer.zero_grad()
        f_l.backward()
        f_optimizer.step()
        # continue

    d_g_l_get =  model(copy.deepcopy(pos_adj),user_batch,rating_batch,item_batch)
    _,f_l,d_l = d_g_l_get 
    loss_current[1].append(d_l.item())  
    f_optimizer.zero_grad()
    d_l.backward()
    f_optimizer.step()

(2) if the prediction is not computed by inner product instead of neural networks(NN), do I need to update the NN together when I update the filters?

Looking forward to your reply!

您好，我注意到原始数据中存在某些缺失值，请问您是使用什么方法填充的呢？非常期待您的回复！