def ctr_eval(sess, model, data, batch_size):
    start = 0
    auc_list = []
    f1_list = []
    while start + batch_size <= data.shape[0]:
        auc, f1 = model.eval(sess, get_feed_dict(model, data, start, start + batch_size))
        auc_list.append(auc)
        f1_list.append(f1)
        start += batch_size
    return float(np.mean(auc_list)), float(np.mean(f1_list))

In this function, you calculate AUC in every batch and take their average value as the final AUC. But as far as I know AUC needs to be globally sorted in the test set. Can you explain it ?

第一列是item的id，第二列是知识图谱中电影的id

Hello Wang,
Book-Crossing 的 item_index2entity_id.txt 和kg.txt 可否给个获取途径，邮箱[email protected]
THanks

如果是增加了新的用户,整个模型是不是只能重新训练???

Hi, Mr. Wang,

I noticed that on main.py there is a block of code to run the program for the Book Crossing dataset which was also mentioned in the paper. However, I think it seems missing some data required to run it. Although I downloaded the Book Crossing dataset from a previous Issue, I could not find other files such as item_index2entitty and kg.txt that are present in both the music and movie modules.

Would it be possible for you to share these files with me, please? Your assistance would be greatly appreciated.

Thank you and best regards,
Luthfi.

Hello Mr or Miss Wang,
In the data folder, why can't I find the data set for the book? And
excuse me, how did you get the data and process it into a usable data set? I want to try to make a new one.

你好，我在写预测模块的时候，发现预测的数据量跟训练模型的batch_size有关，预测的时候喂的数据量必须和模型训练时的batch_size相等才可以预测出结果，我想咨询一下这个是什么原因引起的？

comment the code blocks of parameter settings for MovieLens-20M;

uncomment the code blocks of parameter settings for Last.FM;

请问这两句是什么意思啊

Hi Dr. Wang,

Thanks for sharing the code and data sets for the paper. The movie dataset looks like missing one file. Could you please also share the ratings.csv file for the movie data? Thanks a lot for your help!

Best,
Duna

请问用自己数据构造kg.txt文件时，头实体代表用户，尾实体代表商品，是否用户编号和商品编号不能有重复，不然就无法进行区分？

May I know the version of TF you run experiments on?

self.l2_loss = tf.nn.l2_loss(self.user_emb_matrix) + tf.nn.l2_loss(
self.entity_emb_matrix) + tf.nn.l2_loss(self.relation_emb_matrix)

请问作者有啥可以优化的办法 或者 替换的办法呢？

比如 我的 emb 大小是 [100000000, 32]

OSError: ../data/movie/ratings_final.txt not found.

File "D:\projects\KGCN\src\data_loader.py", line 21, in load_rating
rating_np = np.loadtxt(rating_file + '.txt', dtype=np.int64)

Hello, Professor Wang. Will you use pytorch to implement KGCN in the future?

epoch 29 train auc: 0.9121 f1: 0.8071 eval auc: 0.6914 f1: 0.6629 test auc: 0.7041 f1: 0.6715
①论文里面的指标是用的训练、评估、测试哪个阶段啊？
②上面这个是我用自己的数据跑出来的，请问又有问题吗？

你好，music里面txt里面每个字段具体什么含义呢，例如：2086 music.artist.origin 3846。
我看代码里面对应的是head_old，relation_old，tail_old。看着我很懵，可以做一个比较通俗易懂的解释吗？

Hi, Wang! Should we use tf.reduce_sum to compute user_relation_scores instead of tf.reduce_mean？ Since you said it's a inner product operation in the paper. Thanks in advance.

is it
https://arxiv.org/abs/1905.04413
?

你好作者，kg.txt文件真的不知道怎么创建，我想创建一个ml-100k的kg.txt文件，请帮帮忙

你好，KGCN这个方法相对于其他方法很新颖。我在构建KG图的时候，数据量较小时，处理速度比较快。如果数据量较大时，处理数据的时间就成倍增加。王工对于kg图构建优化上有没有研究？

王教授您好，电影的ratings_final.txt我知道在链接里面，但音乐的ratings_final.txt我找不到，搞了一天到处找，找不到，麻烦您指点一下，辛苦了

Dear Dr. Wang,

Thanks for your wonderful work and clear implementation. I wonder whether you could provide a copy of your processed Book-Crossing dataset with knowledge graph as described in the paper. My e-mail address is [email protected]. Thanks again.

Best

Hi, I am learning knowledge graph based recommender system, and your work is great!
However I have some trouble getting the dataset. For the movielens dataset I know where to download, but for the kg.txt file, I don't know where it comes from. I can't get access to the website https://www.satori.com/. So where do you get the satori knowledge graph dataset?

比如user的数量是item数量的100倍，如果模型performance会收到影响的话，请教一下有什么优化方向吗？还有一个小问题是，训练完之后，直接拿出user embedding和relationship/entity embedding(not for item)做dot product，可以用来解释user的喜好吗？谢谢！！

Hi, i see there us a evaluation called show top, but i'm wondering how can i use the evaluation method?
thank you

The command should be "python preprocess.py -d movie"?

KGNN_LS 的loss 需要引入interaction_hashtable. 这里需要读入全部的训练数据集，如果是训练数据集很大的话，会非常消耗资源（因为全量数据集需要进tensorflow内存）所以问下 这个loss的加入有多大的提升？

Hi, in your paper, you mention that you sample a fixed size set of neighbors for each entity. How do you choose neighbors for each entity? Random or in other ways?

Dear Professor, first of all, thank you for openning source code and data set. In your paper, you mentioned using satori to construct sub-KG, but in the Web page of Bing Satori, I did not find any information about KG, so I'd like to ask how to construct KG from Satori.
Thank you again and look forward to your reply

Hi, could you specify the dependency version (e.g. python3.6 or tf 1.4 etc.) in the repo?

Hello Mr. Wang I'm trying to reproduce your model on a new dataset that uses jurisprudemce documents. My only question is how was the item_index2entity_id.txt dataset generated? Are the item ids the items that appear in the knowledge group only as a tail entity?

Or are the item ids the items that sppear in the knowledge graph as both head and tail entities? For example if a movie item is also a head entity in triple A and is also a tail entity in another triple, say triple B would this be criteria for the item to be part in the item_index2entity_id.txt file or is it enough that an item appear as a tail entity as you've detailed in your MKR paper?

老师好，请问aggregate method里面这段要怎么理解？最后return的entity_vectors[0]是不是之和第一次iteration(hop=0)有关？并且每次iteration里aggregator的input都是相互独立的？

Hi,

I am going through the literature of the paper and one thing which I find missing is information about inductiveness of the proposed algorithm.

So my question is, Is the proposed architecture inductive in nature i.e. it can generalise to new users as well without retraining?

Thanks
Sachin

Hello Wang,
How do we create the knowledge graph file (kg.txt) if we are to train this network on a movieLens 100k dataset?

可以提供一下lastFM生成kg.txt的相关代码吗