alicogintel/sdm

Sequential deep matching model for recommender system at Alibaba

sdm's Issues

num_labels含义

您好，最近在研究您的sdm源码，有一点疑惑：样例数据中num_labels=5，是把每个session最后的5个item作为目标吗？

通过sampled_softmax_loss负采样，inputs=rnn_outputs_split,是用户长短兴趣结合后各个时刻embedding组成的矩阵，那么这边的loss是session中各个时刻都会针对目标num_labels=5进行负采样，然后将所有loss相加？

请问验证时的训练、测试的数据怎么定义？能否给出完整示例？

\SDM-master\code\train\run.py

def run_validating(hparams, model):
acc_keys = ["user_id", "ds", "user_embedding_output"]

#  user defined function
#  you should write your own code here for reading and writing data
**train_file = get_your_train_files()
test_file = get_your_test_files()
writer = open_your_test_result_file()**

不太明白需要什么样的数据格式，是否需要转换，恳请指教。

数据集问题

请问你们京东数据集中各个子段的含义是什么，比如说action.csv中的type的每个取值是什么意思

Datasets

Please upload the datasets as well.

prefer feature choose and tfrecord file generate code

hi, thanks you for open source code. there is two problems

1. is the same prefer feature filtered, such as item_id, brand_id, but not for short feature?
2. Is it convenient to provide the code to generate tfrecord from the original csv file?

label generation logic of JD offline datasets

Hi,fuyu:
Great job for this match framework. However,I'm a little confused about the label generation logic of JD offline datasets as provided in your codes.
If convenient, would you please upload the preprocessing script of the offline dataset. Thx a lot.

Astonishing work! How could you come up with it?

I'm a new one who comes first into the recommender system. I happen to find this work and believe it is excellent and valuable. I want to learn from you and could you impart me some experience?