Question regarding the Pearson correlation calculation

[RemyLau]

Dear author @DHUDBlab, I find your work quite interesting. However, I'm a bit confused about the Pearson correlation computation implemented in your code base:

https://github.com/DHUDBlab/scDSC/blob/1247a63aac17bdfb9cd833e3dbe175c4c92c26be/MTAB/calcu_graph_mtab.py#L29-L32

If I'm not misunderstood, this is inconsistent with the definition of Pearson correlation shown in your paper.

In particular,

1. features are centered by the whole feature matrix. However, if you want the correlation between cells, i.e., rows, they should be row centered.
2. np.linalg.norm taking 2d array as input without any keyword arguments produces the Frobinius norm of the whole matrix. So the denominator part in your computation is inconsistent with the formula provided.

I suggest the following modification, which faithfully implements the intended Pearson correlation calculation:

normalized_features = (features - features.mean(1, keepdims=True)) / features.std(1, keepdims=True)
s = np.dot(normalized_features, normalized_features.T) / normalized_features.shape[1]

You can check that, indeed, the diagonal of s are all ones (with some error due to machine eps).

[lixhere]

hello，could you please send me the datasets?I saw a lot of data on the website they provided, but I don't know which one they used，thx！！！！！！！！1

[handesome]

me too,have you solved the problem?

[handesome]

hello，could you please send me the datasets?I saw a lot of data on the website they provided, but I don't know which one they used，thx！！！！！！！！1

hi!do you konw how to use the data now?could you give some advice?Thanks