oywtece / dstn Goto Github PK

Hi，
Firstly，nice job of you！

However，as mentioned in the paper, there is a tradeoff at serving.

I mean, how do we set the contextual ads if we perform ③ and ④ once？

请问您在这篇论文中对数值特征处理为embeding的时候，用到的quantization以及在论文中提得到的buckets是什么意思？

您在《Deep Spatio-Temporal Neural Networks for Click-Through Rate Prediction》这篇文章中使用了Search advertising dataset和News feed advertising dataset，请问您是怎么处理这些数据集的？ 希望您能在百忙之中抽空解答一下。

For the temporal view in your paper, you use clicked ad sequence and unclicked ad sequence. In your third dataset in Section 3.1, News feed advertising datasets, I think the clicked sequence is ads that a user clicked in the past few days. But for unclicked sequence, do you regard it as a sequence or just sampled impression ads without sequence aspect.

Thanks in advance.

hello，
first of all, nice good job : )
I recently reproduced your paper, but I encountered some problems when using Avito's data set and have not been resolved. So I would like to ask, can you share the code for processing the Avito data set?

老师您好，不好意思打扰到您了。
我想问一个问题，
在config_dstn.py文件中，有一个参数是n_ft = 42301586 # num of unique fts
我没有看懂您给的注释，您在这里定义的n_ft是什么特征的数量？还是有其他含义？

希望能够得到您的解答。

您在论文中只给出了一个数据集的地址，请问另外2个数据集的地址方便给出么？
另外两个数据集是Search advertising dataset 和 News feed advertising dataset

Hi，
Firstly，nice job of you！

What we want to know is what are the corresponding attributes of search_loc_level, search_cate_level, search_sub_cate_id and ad_sub_cate_id in the Avito data file?

抱歉，打扰一下，在阅读论文中思考了这个问题：

我理解context ads是在同一个分页请求结果中在target ad上面的ads。但是这个context ads在模型预测时候如何得到的，毕竟精排只是给每个item依次打分，最终的位次是engine根据所有item打分结果排序得到的，这是果，dstn预测时候应该还拿不到这个果。

谢谢！

Hi，
Firstly，nice job of you！:)

We want to reproduce your experiments on Avito dataset and develop our method on top of DSTN. Could you provide the schema of the CSV files? It would be very helpful to us.

我想请问下，您在训练的时候有没有遇到过tf.exp异常值的问题造成auc跃变，因为din的attention是有softmax，只用tf.exp出现异常值，有没有什么好的避免方法？