stefanfroth/nmf_extension

This repository extends the functionality of the class sklearn.decomposition.NMF

nmf_extension's Issues

Rating values to high

The values of the ratings rise substantially above 5 (eg. 10 or higher). Implement a method to bound the values.

• Initialize so that initial values 0 < iv < sqrt(5/n_components) ?
• Regularization

Add bias

So far it is not possible in the scikit-learn implementation of NMF to add bias terms. Make this possible.

Discussions on sklearn

Hi Stefan (I saw your comment https://github.com/spicedacademy/datascience_source/issues/223#issuecomment-737381613),

for inspiration you maybe also want to have a look at these two issues/ pull requests from the sklearn project:

• scikit-learn/scikit-learn#8447
• scikit-learn/scikit-learn#8474

I think, the idea is, that you modify the loss function in such a way that it ignores missing entries.

reconstruction_err_ not implemented

In the current version of the project the attribute .reconstruction_err_ is not created.

Remove `pandas` dependency

Currently, the code assumes that the input matrix is a pd.DataFrame, because functions such as df.isna() are used. I think one could easily switch to numpy only (e.g. by using np.isnan) which is also the default in sklearn.

Make it faster

The performance of the algorithm is currently very slow. It needs massive speed improvement to be actually useful.

1. Use numpy instead of pandas
2. Use cython

Make Learning Rate adaptive

The learning rate is currently hard coded. It will fail for some problems and work for others. Either adjust the learning rate automatically or introduce it as an attribute.

Allow for NaNs in new vector

Currently the nmf.transform() cannot handle NaNs

Extend README

Include the use of transform in the README file

Prevent item_featue_matrix H from becoming negative

Evaluate Performance

Evaluate whether the performance of this implementation outperforms the original implementation with imputation on the MovieLens Data.