Issues about the torch version code

[YuqiCui]

Hi, I've been read your article and I'm very interested in your work DFKM. I downloaded the USPS data from Kaggle (https://www.kaggle.com/bistaumanga/usps-dataset) and used it to run the torch version of DFKM, but the ACC and NMI seem to decrease during the training. Is there something I need to do with the data preprocessing or change the parameters?

[hyzhang98]

Most datasets are downloaded from http://www.escience.cn/people/fpnie/index.html. It seems like ESCIENCE is unavailable now. I will upload some datasets.

[YuqiCui]

Thank you very much!

[YuqiCui]

Thank you again for sharing the dataset. The code seems to run just fine using the new data. But I can't find any description about the pre-processing in the article. I noticed that this USPS dataset has only 1854 samples but the raw USPS data has 9298 samples. And the range of data is [-1, 1] on the USPS dataset you provided, and [0, 1] on the raw USPS dataset. Could you please tell me how did you downsample this dataset and how to pre-processing?

[hyzhang98]

I notice that all datasets downloaded from Prof. Nie's homepage are preprocessed in 2010, and they are widely used in many clustering/dimensionality-reduction papers. I tried my best to find the details, but the paper that uses this subset is really hard to find. If you are still interested in the details, I think you can look up the clustering-related papers of Prof. Nie (google scholar). And if I find the corresponding description, I will inform you at Github.

[hyzhang98]

You can use email to contact. My email: hyzhang98@gmail.com or hyzhang98@mail.nwpu.edu.cn. I will regularly check it.

[zhangyue233]

ok, thanks