cheungdaven / deeprec Goto Github PK

hi,
In the evaluation process of AttRec, when calculating the recommendation item rank of each user, why use the negative items instead of all items?
this code in: DeepRec/utils/evaluation/SeqRecMetrics.py--->def evaluate1(self)

In my understanding, It should be to calculate the score of each user and all items, then rank them, and finally take the top 50 to calculate HT@50.

Calculate it the way the code says, it does not have any positive items in the prediction process.

i have 2M user and 300k items, ram is not enough, How do you implement????

Hi,

If I wanna feed more features to the models shall I only change the headers list? can I pass a vector to it as well? Many thanks

In practice, the number of users or items are usually extremely huge, So, tf.nn.embedding_lookup seems operation is impossible. How to deal with it?

HR@50:0.3753976670201485
MRR:0.053446402676701824
AUC:0.8771433443062608
92456
Epoch: 0543; (943, 100)
how can I reproduce effect

The first output is

Epoch: 0005;
HR@50:0.5005302226935313
MRR:0.08745163611856224
AUC:0.9075360408864436
2020-02-07 09:50:03

then

Epoch: 0123;
HR@50:0.46977730646871685
MRR:0.08698753010475276
AUC:0.8941231973552161
2020-02-07 10:00:21

Aren't these two bigger the better????

您好，有幸拜读您的这篇文章”Next Item Recommendation with Self-Atention”。所以想问一下您有没有这篇论文的代码，作为刚入学推荐系统的小白，还是希望再拜读一下您的代码，非常感谢！（知道您是**人，就不写英语了哈哈）

for line in test_data.itertuples():
    test_row.append(line[1] - 1)
    test_col.append(line[2] - 1)
    test_rating.append(1)

Why line[1] and line[2] add the -1?

According to the codes, this function evaluate() generates scores for negative items and then rank and get top k items for later evaluation.

However, in map_mrr_ndcg() and precision_recall_ndcg_at_k(), variable hits is calculated by seeing if any single negative item is in test data. If the given negative item set does not contain test item, hits will be 0. This is ridiculous and significantly affects the evaluation results, especially when taking out only 100 negative samples randomly.

The performance of a model highly depends on if it is lucky enough that its negative samples have test set.

In file testSeqRec.py I find this piece of code:

        if args.model == "AttRec":
            train_data = DataSet(path="../Data/ml100k/temp/train.dat", sep="\t",header=['user', 'item', 'rating', 'time'],isTrain=True, seq_len=5, target_len=3, num_users=943, num_items=1682)
            test_data = DataSet(path="../Data/ml100k/temp/test.dat", sep="\t", header=['user', 'item', 'rating', 'time'], user_map=train_data.user_map, item_map=train_data.item_map)

but I can't find ml100k/temp/train.dat and ml100k/temp/test.dat, there is only movielens_100k.dat file in the ml100k/ folder, could you please provide the data processing procedure?

Hi,
I find There is a wrong annotation for your code NRR and I can't find the right paper for this method.

As a recommender system application developer, I find your DeepRec is really helpful, I just download your source code and study it. Super duper thanks!

I am wondering if possible to add HRNN-Meta implementation inside your project?
HRNN-Meta is created by Amazon, you can read the paper here:
https://openreview.net/pdf?id=ByzxsrrkJ4

Amazon also uses HRNN-Meta created a SaaS service named Personalize:
https://aws.amazon.com/personalize/
The algorithm is inspiring, and the result also looks very impressive.

您好，有幸看到你的代码，请问这个代码放到eclipse应该怎么运行

原文中的损失函数如下：

与代码中实现的不一样。

对照了BPR的论文和代码，是不是这里的实现是在caser原文似然公式后也加入了参数\theta的先验分布（高斯分布）？

不知道理解的对不对。

How do i run this code on Jupyter notebook

作者您好，
因为自动编码机那地方，是评分矩阵输入到隐藏层编码完成，隐藏层到输出层就得到预测评分矩阵。
请问作者，这个代码的意思是对评分矩阵进行编码，然后解码，从而重构并补全评分矩阵吗？

作者您好，
我想问下您的基于项目的编码协同过滤中训练之后，在哪里保存的训练参数呢？还是说系统自动保存？

1. 代码实现中num_hidden没用上
2. 看起来是把特征嵌入的维度与注意力嵌入的向量维度都设置为num_factor=16了
   我理解num_hidden是特征嵌入维度参数，num_factor是注意力嵌入维度参数吧？

I find the data in "test_item_ranking.py" is still in the range of 1-5, should it be 0/1 in the ranking problem? For example in mlp and GMF model

Hi, quick question: do you have any benchmark available on a common dataset that validates the correct implementation and performance of the models?

您好！最近细读了您的论文：AutoSVD++: An Efficient Hybrid Collaborative Filtering Model
via Contractive Auto-encoders。

想对其中的自编码机处理content information的过程深入了解。在您的github没有找到处理item_feature的相关程序。本人水平有限，对content information的处理没能力复现。
请问您方便指点下content information的处理细节吗
（from 'item side information' to 'latent vector' of item）

无法运行，一些函数调用都有问题；
例如 函数lram 中的dropout参数