cosanlab / neighbors Goto Github PK

We should add sphinx documentation similar to nltools package. Readthedocs seems to work well. @infiniteline interested in taking a crack at this?

Should cast this line using nan to num: https://github.com/cosanlab/neighbors/blob/master/neighbors/_fit.py#L71

We should verify that we are actually centering the kernel convolution as is mentioned in the documentation, or if instead we are doing a forward convolution (which is what we are likely doing). This is a very small potential bug that I doubt will amount to any appreciable difference, but would good to ensure that we are doing what we say.

This would ideally be in base class.

This would be a good idea to add at some point.

We should add sphinx documentation similar to nltools package. @infiniteline interested in taking a crack at this?

Might only be working with number of iterations.

At some point we need to add an extension to accommodate tensors for multivariate rating data. Not sure how difficult it will be.

uses alternating least squares optimization

Currently we force a boxcar, but it would be a trivial extension to allow any kernel shape to be passed into the convolution.

Error with missing key 'user' at this line: https://github.com/cosanlab/neighbors/blob/master/neighbors/utils.py#L620

right now vmin, vmax for heatmap is fixed to 0,100

I am running estimate_performance() function on sparse data and am curious about how the mask_items keyword argument works.

From docstrings
n_mask_items (int/float, optional): how much randomly sparsify dense data each iteration; Defaults to masking out 20% of observed

This keyword makes sense for dense data, but how does it work with sparse data? is it ignored?

This will dramatically impact results. Correct solution is to update mask after dilation to ensure that a bunch of zeros aren't added to data.

from neighbors.models import NNMF_sgd
from neighbors.utils create_user_item_matrix, estimate_performance

In your example code in the README, the second import statement is missing the word 'import'.

Everything else seems to be working well.

This results a "fragmented dataframe" warning. Could refactor the lines around here: https://github.com/cosanlab/neighbors/blob/master/neighbors/utils.py#L283

change:
ratings = pd.DataFrame(columns=df.Item.unique(),index=df['Subject'].unique())

to:

columnNames = sorted(df.Item.unique())
ratings = pd.DataFrame(columns=columnNames,index=df['Subject'].unique())

Could potentially add joblib and parallelize iterations.