pharo-ai/k-means

K-means clustering in Pharo

k-means's Issues

Prefix classes with ML?

I would like to propose to prefix the KMean classes with ML, as there is another implementation of KMeans which uses the same name, and it would be more consistent with new packages like MLViz and MLMetrics.

What do you think?

Add parameter to run the k-means algo several times and return the best found cluster

https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html

n_initint, default=10

Number of time the k-means algorithm will be run with different centroid seeds. The final results will be the best output of n_init consecutive runs in terms of inertia.

Add max iterations for the case in which the algorithm takes a lot to converge

Now we are only checking if the algo is converging. We should add a max iterations parameter that is initialized with a number by default.

Implement coverage method

We should have a method, common to all algos, that returns a boolean saying that the algo has reached convergence or not. We need to separete that of the max_iterations.

If the algo has reached the max iterations and has not converged it should return false.

This should be for all algos

Rename package and baseline to follow ML*

Do we want to have AI for non ML one?
But in any case we should have AI or ML.

Add class comment!!!

Add transform: method

The method transform will return the distance of a variable to all the centroids. It is useful for dimensionality reduction.

More info: https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html#sklearn.cluster.KMeans.transform

You should be able to define initial centroids

When initializating kmeans, you should be able to choose if you want the initial centroids to be random or #6, or if you want to set your own initial centroids. Keep in mind that If an array is passed, it should be of shape (n_clusters, n_features) and gives the initial centers or else raise an exception.
https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html.

Adding this feature would be helpful in testing that empty clusters are relocated as expected if you give one initial centroid that is far from the data which means that a cluster will be empty on the first iteration

Add hard clustering and soft clustering methods

Add tests

Double dispatch for #fit: method

Currently I'm using k-means with DataFrame as follows:

| df kmeans |
df := Datasets loadIris columnsFrom: 1 to: 4.
kmeans := KMeans numberOfClusters: 3.
kmeans fit: (df asArrayOfRows collect: #asArray).

It would be nice if the argument to #fit: could be just the DataFrame, which knows how to be fitted with a KMeans algorithm:

kmeans fit: df

This way fit could receive also PMMatrix and similar matrix-like objects. And each one is responsible to implement:

DataFrame>>fitKMeans: aKmeans
  aKmeans fit: (self asArrayOfRows collect: #asArray).

and so on...

nbClusters should be between 1 and data size

This

 | data kmeans |
	data := #( #( 0 0 ) #( 0.5 0 ) #( 0.5 1 ) #( 1 1 ) ).

	kmeans := AIKMeans numberOfClusters: 100.

	kmeans fit: data

and this

| data kmeans |
	data := #( #( 0 0 ) #( 0.5 0 ) #( 0.5 1 ) #( 1 1 ) ).
	kmeans := AIKMeans numberOfClusters: -100.
	kmeans fit: data

should raise an appropriate exception because now the first one either works or says Tooktoomuchtime while the second raises a SubscriptionOutOfBound

Add inertia parameter

inertia_float

Sum of squared distances of samples to their closest cluster center, weighted by the sample weights if provided.

https://scikit-learn.org/stable/modules/generated/sklearn.cluster.KMeans.html

Implement k-means++ centroids choosing algo

https://scikit-learn.org/stable/modules/clustering.html#k-means

Now, the centroids are selected randomly. We can do better than that using the k-means++ like in scikit-learn.