hinanmu / mlknn Goto Github PK

Hi,
I'm running your code and find that the performance (i.e. hamming loss) is slightly worse than that stated in the paper.
By checking the code I find some, I think, bugs there:

1. wrong length of label vector: the original label data provided in dataset/ are 13 in length, but it should be 14. I find this is raised by the wrong starting point of slicing in line 22 & 24 in data_process.py.
2. mistaken assignment of Peh0/1: in line 126 & 127 of mlknn.py, Peh0 & Peh1 are assigned as if they are scalars, but they are actually matrices. I guess they should be indexed by [i][j] before being assigned.
3. failed loading: in the load(·) function of MLKNN, the loaded data are assigned to some non-member variables. I guess there should be a prefix self. before those l-values.

But despite fixing all these, I get a similar result (0.3037077426390403) as before (0.30307860078852444).
I also find that result remains unchanged after Ph0/1 & Peh0/1 being set to all zeros in the front of the test(·) function, but becomes worse if they are set to random values.
Any thought?
Thanks