NaN matrix produced by TCA python code about transferlearning HOT 6 CLOSED

@zzbslayer I know what happened. It's not the problem of too small values, it's the divide by 0 issue during the normalization of X and Z (basically your data contains too many 0s):

Line 64: X = np.dot(X, np.diag(1 / (np.sum(X ** 2, axis=0) ** 0.5)))
Line 78 Z = np.dot(Z, np.diag(1 / (np.sum(Z ** 2, axis=0) ** 0.5)))

If you uncomment these two lines, TCA will work.
You can replace these two lines with other normalization methods.

from transferlearning.

from transferlearning.

@zzbslayer Could you please send me your data? It's better in the Xs, Ys, Xt, Yt format

from transferlearning.

Sure. Let me do some process. The original dataset is a bit confusing and difficult to use.

from transferlearning.

transfer.zip

@jindongwang Here're the Xs and Xt inside the zip file.
They can be accessed by the following code. Ys and Yt seem unnecessary using TCA.

Xs = pickle.load(open("Xs.pkl", 'rb'))  // np.array
Xt = pickle.load(open("Xt.pkl", 'rb'))  // np.array

from transferlearning.

@jindongwang Oh so surprising.. I really appreciate it !

from transferlearning.