about the missing rate

[iTomxy]

Hi, I wonder, why the missing rate is halved here?

[Lin-Yijie]

Thanks for your interest in our work! It is just a showcase. Our method could handle the incomplete problem of different missing rates and you can change it in configure.py.

[iTomxy]

As in the code, the missing rate is seemed to be determined in the get_mask function. However, if I comment out this line, or change the missing rate to 1 in configure.py, the code will fail to reproduce the results under the incomplete setting (i.e. missing rate = 0.5).

[Lin-Yijie]

The masked data is generated by the get_mask function and the corresponding missing rate is in configure.py. It should be pointed that the missing rate in configure.py is the real missing rate as talked about in the main paper. For the line missing_rate = missing_rate / 2, it is the process of generating incomplete multi-view datasets. By the way, our method is trained on the complete multi-view data thus cannot handle the situation with missing rate=1.

[iTomxy]

Well, you're right as I compute the missing from the resultant mask and it's around 0.5. I guess you did that halving because there are 2 views and you want them each bear 0.25 missing rate. And I suggest using:

missing_rate = missing_rate / view_num

which may be clearer.

[Lin-Yijie]

Thanks for your suggestions! I have changed the code to missing_rate = missing_rate / view_num now.

[zhangyuanyang21]

Hi, I would like to ask if running a dataset with the missing rate set to 0, is it a complete view network model?

[Lin-Yijie]

Hi, I would like to ask if running a dataset with the missing rate set to 0, is it a complete view network model?

Yes it is~

[zhangyuanyang21]

Thank you very much for your reply,

[zhangyuanyang21]

你好，我发现Competer网络只能跑两个视图的数据集超过两个视图就不可以了，请问是不是这样的呢？

[Lin-Yijie]

Exactly. You could refer to our multi-view version DCP.

[zhangyuanyang21]

Thank you very much for your reply！