A large portion of the Python user base doesn't use Python 2 any more, and therefore can't use a package that doesn't support Python 3. Please add Python 3 support; it shouldn't take too much extra work.

I already install skfeature-chappers (1.0.2) ,but i got an AttributeError.

AttributeError: module 'skfeature.function.similarity_based.reliefF' has no attribute 'feature_ranking'

Hey! Could you please provide an example code for calculating fisher score present in the path

skfeature.function.similarity_based.fisher_score

Could you please help me with what class labels have to be provided as the function should extract the fisher features and provide the labels.

Using concurrent futures, which requires Python 3.2 and newer (released February 20th, 2011).

Tested on Windows Server 2016 DataCenter 64 bit, with Python 3.6 using the Conda Python distribution.

Note that on Windows you need to have in the calling code: if __name__ == '__main__': This is a general Python on windows issue with spawning new processes, but including it does not stop it working elsewhere.

This parallel version gives consistent results to the previous version, in my testing.

Hello:

I am using the unsupervised feature selection using Laplacian Score. But I am facing the below error message

Traceback (most recent call last): File "fs.py", line 12, in W = construct_W.construct_W(frame, *_kwargs_W) File "/usr/local/lib/python2.7/dist-packages/skfeature/utility/construct_W.py", line 141, in construct_W D = pairwise_distances(X) File "/usr/local/lib/python2.7/dist-packages/sklearn/metrics/pairwise.py", line 1207, in pairwise_distances return _parallel_pairwise(X, Y, func, n_jobs, *_kwds) File "/usr/local/lib/python2.7/dist-packages/sklearn/metrics/pairwise.py", line 1054, in _parallel_pairwise return func(X, Y, **kwds) File "/usr/local/lib/python2.7/dist-packages/sklearn/metrics/pairwise.py", line 231, in euclidean_distances distances = safe_sparse_dot(X, Y.T, dense_output=True) File "/usr/local/lib/python2.7/dist-packages/sklearn/utils/extmath.py", line 184, in safe_sparse_dot return fast_dot(a, b) MemoryError

I have updated the scikit-learn; but the issue persists. Any inputs will be helpful

Regards Sayantan Guha

In the fisher score calculation (https://github.com/jundongl/scikit-feature/blob/48cffad4e88ff4b9d2f1c7baffb314d1b3303792/skfeature/function/similarity_based/fisher_score.py) , it was mentioned L = D - W (line 10) but in the code, L was simply defined as L = W (line 38)

I'm not sure is this a bug or if I miss out anything? Anyone came across the same issue?

Hi, thanks for your great work. And I run into an error when I try to run the test_MRMR.py example. It seems like some code are on the Python2.x, but when try to run on the Python3.x, it crashed.

Could you please update these part of code, make it fit in Python3? Thx a lot.

According to reference paper, 'Unsupervised Feature Selection for Multi-Cluster Data' by Cai,Deng, in equation 4, max of mod value is taken. I think that needs to be corrected here in MCFS.py in line 69 from W.max(1) to np.absolute(W).max(1)

D:\prj>python findFeatures.py Traceback (most recent call last): File "findFeatures.py", line 7, in from skfeature.function.information_theoretical_based import MIM # infogain File "D:\ProgramData\Anaconda3\lib\site-packages\skfeature\function\information_theoretical_based\MIM.py", line 1, in import LCSI ModuleNotFoundError: No module named 'LCSI'

I am using this library for my research and I noticed some instability with the algorithm.

I then checked the code and it seemed like the algorithm was using random sampling with duplicates (when iterating over n_samples anyway...).

So I made a change and a pull request. The detailed explanation is below:

The previous version used random sampling with duplicate examples from the dataset. The algorithm idea states that it should use random sampling of m examples without duplicates because duplicates cause unstable results where the effect of some examples will be aplified and some will be ignored. Whereas since the current code set m to n_samples (all examples), there's really no point in random sampling since you use all examples anyway. The order of sample processing doesn't matter either, so I removed all this random sampling bit and just let it loop through results.

In future I can modify the algorithm to use another argument m and use a random subset of examples properly (like stated in the article), but for now I did it like this.

The algorithm is stable now. It might be cook to have this sorted on master as well since some people are using it and it doesn't perform too well now.

I can also do the m samples subset modification later on. Let me know.

Sincerely, Tadej Magajna

corpus, categories = get_detail_content_category()

vectorizer = TfidfVectorizer(max_df=1.0, max_features=6000, min_df=1,
                             stop_words=get_cn_stopwords(),
                             encoding='utf-8', decode_error='ignore',
                             analyzer='word', tokenizer=cn_tokenize)

X = vectorizer.fit_transform(corpus)
y = categories

idx = ICAP.icap(X, y, n_selected_features=1000)
selected_X = X[:, idx[0:1000]]

After I run this code, i get error like title. I don't why, any help is appreciable.

Around line 39 in the code for FCBF uses np.zeros((n_features, 2), dtypes='object') which throws an error. Numpy documentation states the parameter should be dtype (without the 's'). So the line should be np.zeros((n_features, 2), dtype='object') instead. It was an easy enough fix on my own and I saved a separate .py file with the fixed version and it worked.

The original function involves a lot of repeated operations, and when the number of features is large, it takes a lot of time. Using a dict R to store the calculated correlation values can improve the speed by 1-2 orders of magnitude.

In https://github.com/jundongl/scikit-feature/blob/master/skfeature/function/similarity_based/reliefF.py

In the Algorithm Relief-F

（ This figure is from https://link.springer.com/article/10.1023/A:1025667309714 ）

The equation P(C)/(1-P(Class(Ri))) is in the in the numerator

The variable corresponding to this equation in the code is p_dict as below

        p_dict = dict()
        p_label_idx = float(len(y[y == y[idx]]))/float(n_samples)

        for label in c:
            p_label_c = float(len(y[y == label]))/float(n_samples)
            p_dict[label] = p_label_c/(1-p_label_idx)
            near_miss[label] = []

So in the code of calculate weight, p_dict should be in the numerator and not the denominator.

The code of

            score += near_miss_term[label]/(k*p_dict[label])

should be

            score += (near_miss_term[label] * p_dict[label]) / k

Based on the code implemented in the calculation of trace ratio, if we set s_within and s_between to have a length of n_selected_features. Then the variables "I" and "idx" will just have a length of n_selected_features. In this case, arent we just restricting the selected features to be the first n_selected_features in fs_idx and the loop is essentially updating/optimizing nothing?

# preprocessing
fs_idx = np.argsort(np.divide(s_between, s_within), 0)[::-1]
k = np.sum(s_between[0:n_selected_features])/np.sum(s_within[0:n_selected_features])
s_within = s_within[fs_idx[0:n_selected_features]]
s_between = s_between[fs_idx[0:n_selected_features]]

# iterate util converge
count = 0
while True:
    score = np.sort(s_between-k*s_within)[::-1]
    I = np.argsort(s_between-k*s_within)[::-1]
    idx = I[0:n_selected_features]
    old_k = k
    k = np.sum(s_between[idx])/np.sum(s_within[idx])
    if verbose:
        print('obj at iter {0}: {1}'.format(count+1, k))
    count += 1
    if abs(k - old_k) < 1e-3:
        break

UDFS.py L97: The additive parameter \lambda should be independent from gamma (introduced in eq 8 in the paper). It should probably default to something small. It's just used to make the covariance invertible.

Also, construction of S_i seems incorrect to me: UDFS.py L100: indexing on idx_new should be idx_new[:,q]?