Functions and utilities

[verri]

Let's discuss here which functions the package should provide.

Based on the paper, I think the following are a must-have:

• sledge_descriptors(X, labels, minimum_support=0): given a data frame X, where each column corresponds to a pattern and each value x_ij the presence/absence of the pattern j in the sample i, and the clustering results labels, this function returns a list of sets, each of them containing the 1-itemsets that describe a cluster (after particularization.)
• sledge_score(X, labels, minimum_support=0): same inputs, returns a list containing the SLEDge score for each cluster.

Another optional function that I think of, is something like sledge_matrix that returns the values of S, L, E, and D for each cluster (that is, before aggregation with the harmonic mean.) However, this functionality can be incorporated in sledge_score with some trigger.

[verri]

Since, conceptually, sledge_score would call sledge_descriptors, it also makes sense to provide a variation of the sledge_score function that has as arguments the output of sledge_descriptors. Maybe something like:

• sledge_descriptors(X, labels, ...) -> [ [...], [...], ... ]
• sledge_score_from_descriptors([ [...], [...], ... ], ...) -> [ ... ]
• sledge_score(X, labels, ...) == sledge_score_from_descriptors(sledge_descriptors(X, labels, ...))

[verri]

Moreover, given the desired smooth integration with sklearn, while implementing/documenting we should expect arguments compatible with the outputs of the clustering techniques implemented there. Also, it is wise to use similar parameter names.

[verri]

I think we should provide at least auxiliary functions to plot the “SLEDge curve” for a given cluster.

I vote for producing some output that is easily fed by matplotlib. Then, we write some tutorials to create the plot itself (like sklearn does.)

[verri]

Taking a look at sklearn's documentation, I suggest a minor change:

• sledge_descriptors: returns a sparse support matrix for clusters vs patterns.
• sledge_clusters: returns the SLEDge score for each cluster.
• sledge_score: returns the average (or min, or max too?) SLEDge score.
• sledge_curve: returns the SLEDge curve.

Rationale:

• All *_score functions in sklearn return a single value.
• sklearn.metrics.silhouette_samples returns a silhouette score per sample. It makes sense to provide a SLEDge variation per cluster with similar naming.
• There are some *_curve functions in sklearn.metrics.
• All sklearn.metrics.*_matrix functions return a square matrix, so although sledge_descriptors returns a matrix, the name descriptors is more informative and may avoid misinterpretations.

[verri]

Even better:

• sledge.sematic_descriptors
• sledge.sledge_score_clusters
• sledge.sledge_score
• sledge.sledge_curve