While I was looking through the code, I noticed that in neigh_samplers.py, users can sample themselves as the second-order neighbors because when we sample 10 neighbors of the first-order neighbor, the user is also included.

        adj = self.adj_info[node, :]
        neighbors = []
        for neighbor in adj:
            if first_or_second == 'second':
                if self.visible_time[neighbor] <= timeid:
                    neighbors.append(neighbor)    

I was wondering is this intended or a logic flaw. Thank you!

After running the model, I am wondering if weights of attention and their visualization can be done as the attached file displays from the original paper.

There are many examples for extracting and visualization for seq2seq but couldn't really find one for feature explanation.

Is there any ideas/code that can be used for attention visualization of meaningful features? ?

I am working on my own data -purchase history of financial production. So I have data as who, with features, purchased which products, with features and when. However, I have found that the metric of the model is AUC -- and I, as expected, get the following error

ValueError: Only one class present in y_true. ROC AUC score is not defined in that case.

Since all my Ys are 1s -as I only have who purchased (clicked) which and there is no 0 value ( who didn't purchase which). As other CTR prediction models also use AUC and log loss, I assume that there must be a way to use AUC for such a dataset.

Do you have any idea how to solve this issue? That would be really helpful!

I am very impressed about the usefulness of AutoInt to my personal project. I'm wondering if there is any handy way of converting this AutoInt to top-n recommendaiton? Would it be possible?

@Songweiping ,Hello,I am very interested in your work(AutoInt) at CIKM'19,but I had some doubts when I reproduced the experiments of the paper. As we know, Criteo has its own bench mark, including two csv files(train.csv & test.csv).But the preprocess of AutoInt splits the file(train_samples.txt) to get the train data,valid data and test data. I'm wondering how to transform the original Dataset(train.csv & test.csv) into a Dataset that the data_process code can handle. Could you post your code for converting original csv files(train.csv & test.csv) into the whole txt file(the data used in your paper, and Format is similar to the sample file train_example.txt)?

In multi-head attention, there is a relu after queries, keys, and values. Is this a correct implementation? The paper did not mention the relu in Eq. 5. Besides, it seems that the relu will make the attention matrix always positive.

# Linear projections
Q = tf.layers.dense(queries, num_units, activation=tf.nn.relu)
K = tf.layers.dense(keys, num_units, activation=tf.nn.relu)
V = tf.layers.dense(values, num_units, activation=tf.nn.relu)```

在SocialRec中，samples1和samples2的默认值分别为10和5。如果我没理解错的话samples1代表第二层邻居的数量，samples2代表第一层邻居的数量，这样的话第二层邻居的数量就比第一层邻居的数量还多，与经验不符。而且您的论文中也写到“The neighborhood sample sizes are empirically set to 10 and 15 in the first and second convolutional layers”，请问是少写了一个1吗？

Hi, I'm trying to test the result in the paper Session-Based Social Recommendation on the Delicious datasets . From the paper and datasets, I understand that you consider each session is a sequence of tags a user has assigned to a bookmark and all the tagging actions for that bookmark will have the same timestamp according to the datasets. So my question is that when you give the time_id for each session (as in the preprocess_DoubanMovie.py file), will two sessions have the same id only if their timestamps have exactly same date and time ? E.g: A session with timestamp '01/06/2020 2:30:00 pm' and another session with timestamp '01/06/2020 2:30:01 pm' will have different time_id.