xfurna / coalapy Goto Github PK

Quoting myself-

I was experimenting around with some omics data taken from TCGA and I came across cases where multiple modalities had the same relevance. The paper has not discussed such cases. I thought dampening the weightage of modalities with the same relevance might evaporate the cluster information. And when I ran the algorithm on such data, keeping the value of alpha for the modalities with the same relevance equal (phrasing alternatively: if I keep 阝= 1), I found that the cluster labels were biased towards the clusters of that modality which occurred prior to others.
So for example, if I had two modalities with equal relevance then my final cluster labels would have more common labels with the ground truth of that modality which was fed to the algorithm as X1.

Refer the following pseudocode-

# column normalization
            for(i in columns(V))){
                rms<-rootmeansquare(V)
               V[:,i]/=rms
            }
# row normalization
            V=rownormalize(Lkstar)

Scope of improvements in relevance computation.
Try-

def compurte_alpha(...):
    .
    .
    .
    return your_alpha
your_alpha=compute_alpha(...)
new_alpha = [Alpha/np.sum(your_alpha) for Alpha in your_alpha]
.
.
.

Gaussian kernel has issues. Look into it.

The orthonormal basis is taking low rank approximation of laplacians. Which seems off-track. Look into it.

Mix Central Limit Theorem with your data.
try manual scaling of data points-

(x - mean(x)) / sd(x)

Write tests.

Some data may need to undergo a transformation before processing (like power, log, etc..).

Orthogonalisation process giving wrong (though orthogonal) results

It goes with #15 as eigenvector under consideration has not been omitted.

recommended: 1e-10

np.partition is applied incorrectly. Fix it.

Try apply this for it.

Every feature in each modality can be weighted according to the structure it provides to the clusters.

• Error while running tests on RawToLap with generate_data within it.
• generate_data alone however is cool.
• verbose-

(base)  evi1haxor@Devi1ixir > /hdd/Ztudy/BTP/code/CoALa/algo >  master ● > python checks.py
Saving toy data csv w/d 3
Saving toy data csv w/d 3
recieved dframe_csv object with path- /hdd/Ztudy/BTP/code/CoALa/algo/X1.csv
calling matrix with df_csv args /hdd/Ztudy/BTP/code/CoALa/algo/X1.csv
Traceback (most recent call last):
  File "checks.py", line 69, in <module>
    RawToLap()
  File "checks.py", line 58, in RawToLap
    x1 = src.modalities.modality(path_x1, mat_type="gaussian")
  File "/hdd/Ztudy/BTP/code/CoALa/algo/src/modalities.py", line 8, in __init__
    self.W = helperFunc.get_similarity(dfhandler.dframe_csv(path, mat_type=mat_type))
  File "/hdd/Ztudy/BTP/code/CoALa/algo/src/helperFunc.py", line 13, in get_similarity
    mat = matrix(df_csv = df_csv) #dframe_csv obj
  File "/hdd/Ztudy/BTP/code/CoALa/algo/src/helperFunc.py", line 23, in matrix
    return sm.Gaussian(df_csv.df, df_csv.df.shape[1])
  File "/hdd/Ztudy/BTP/code/CoALa/algo/src/matrices.py", line 12, in Gaussian
    A[x][y]=hf.dist(x+1,y+1, df)
  File "/hdd/Ztudy/BTP/code/CoALa/algo/src/helperFunc.py", line 35, in dist
    result_vector = df[list(df.columns)[i]] - df[list(df.columns)[j]]
IndexError: list index out of range

It seems omitting first eigenvector would yield better results.

Laplacian seem to have varied precisions. Round them off once and for all.

Test dataset has modalities with equal relevance. So due to assignment of inappropriate alpha, tests are failing.

commit failed tests. Maybe test script needs to be updated.

The resulting orthonorm_basis from either package are not aligning beyong rth column.