optimum-selector

“Optimize” a matrix by successively removing row-column pairs.

For now, this process only works on square matrices. Each iteration removes a row-column pair as specified by a given score function. The score function takes the current matrix as input and needs to return the next row-column index that should be removed.

In addition, a summary function can be specified, which takes the whole matrix as input and computes some value (e.g., mean of all cells).

The selector uses masked arrays, i.e., no data is actually removed. Therefore, the indices remain stable during processing as the array shape stays the same.

The main script (optimum_selector.py) contains some example score functions. For example, min_mean_col removes the row-column pair with the smallest mean column value.

The Selector class expects a numpy.ndarray as input. The main script currently only supports loading gzip-compressed CSV data that was written with NumPy‘s savetxt function.

Usage

The input needs to be a square 2D matrix supplied in the CSV format with a header row. The header row should start with # since the data is loaded with NumPy‘s loadtxt function. The headers are used to label the removed row-column pair in addition to the index.

Example input

# 123, 345, 567
   0 ,  1 ,  2
   3 ,  0 ,  4
   5 ,  6 ,  0

Score Function

There currently is no elegant way to select the score function. The desired function needs to be specified in the Selector constructor:

selector = Selector(headers,
                    data,
                    score=max_weighted_dist, <- Here
                    summary=np.mean,
                    mask_value=args.mask_value)

Example Call

python3 input_data.csv.gz output_data.csv.gz -m 0

The output is written to a gzip-compressed CSV file. The -m parameter can be used to mask (ignore) an arbitrary value (0 in this example) from the input data and all calculations.

Output Format

The output file contains three columns summary,label,idx, where summary is the output of the specified summary function (if applicable) for the current matrix, label is the label of the removed row-column pair, and idx is the index.