weitianxin / macr Goto Github PK

Hi, I'm still a bit confused about how you configured the valid set.
To construct a valid set from the data you provide, should it be created in the test file or in the training file?

Also, I ran the code using gowalla dataset, but I achieved higher recall (0.10) than the value you mentioned in the paper. I'm wondering why the results are different.

Thanks for your help.

Hi!
I wonder whether the valid set is included in test.txt.
In the paper, you mentioned that test data takes 10% of the entire data, but it seems like the data you've uploaded has more interactions than that. (For instance, gowalla has 1027370 interactions in total, and test data has 204905 interactions which is around 20% of the entire dataset)

If not, I would like to ask if I can get a code for the valid split.

Thanks!

for example, figure1.
thanks a lot.

Hi,
I would like to know if this article uses an early stop strategy? It seems that training 1000 epochs would take a lot of time.

Thanks!

Your paper is very helpful to me as a beginner who is interested in causal recommendation! But when I read the code, I encountered a little confusion. May I ask why the BCE loss function in the code considers positive and negative samples? Like the following code, it seems to be inconsistent with the formula (6) given in the paper?

self.mf_loss_ori = tf.reduce_mean(tf.negative(tf.log(tf.nn.sigmoid(pos_scores)+1e-10))+tf.negative(tf.log(1-tf.nn.sigmoid(neg_scores)+1e-10)))