@solzimer are you planning to add any algorithm to define the initial clusters, besides the random approach?
It seems the K-means++ algo is a good one, which speed up the calculation and increase the probabilities to reach a global-optimum.

When using custom distance function, the passed centroid to the function is undefined. The index for getting the centroid should be j instead of k.

Sometimes two data points should never be put in the same cluster, I'd like to assign them a high distance value in those cases, which I believe would require a custom function. Ideally something where you could return 1.0 for "big distance" and then call the super method for the real distance in the rest of the cases.

Hi, I was thinking of displaying this data maybe using plotly
https://codepen.io/plotly/pen/ojVoQY
Have you implemented a visualisation yourself? Which tool did you use and how did you format the data into x, y, z prior to visualising?

Vince

The function:

// kinit.js
kmpp(data,k) {
  var distance = data[0].length? eudist : dist;
  var ks = [], len = data.length;
  var multi = data[0].length>0;
  var map = {};
// ...
}

Does not use the user-provided distance function. Something like this would remedy:

// kinit.js
kmpp(data,k,fndist) {
  var distance = fndist? fndist : data[0].length? eudist : dist;
// ...

// main.js
else if(initial=="kmpp") {
  ks = kmpp(data,k,fndist);
}
// ...

@solzimer I used this dataset (500 random points) for this simple test:

var result = skmeans(dataset(), 16);
console.log(result.centroids);

However I noticed sometimes one or two centroid are set to [0, 0], which is completely out of range (here a graphical representation of the result).
I tried to debug the code to trace the issue (I believe it's a bug), but I couldn't find any useful hint.

Could you please check this out?
Due to the indeterminability of the output, unfortunately I can't give you a more specific example to better highlight the issue, but please let me know if I can be of help.

There doesn't seem to be any license attributed to this repo. Please clarify the licensing of this library. BTW, thank you @solzimer for this super simple implementation!

This might be a stupid question.

I was reading https://adamspannbauer.github.io/2018/03/02/app-icon-dominant-colors/ and stumbled across this library.

The code on that blog article that I'm questioning is:

    #cluster and assign labels to the pixels 
    clt = KMeans(n_clusters = k)
    labels = clt.fit_predict(image)

    #count labels to find most popular
    label_counts = Counter(labels)

    #subset out most popular centroid
    dominant_color = clt.cluster_centers_[label_counts.most_common(1)[0][0]]

Is it possible to see which centroid cluster is the most common, or is that not really part of the k-means method?

On this line, a data point is assigned to the centroids array by reference:

This means that as the centroids change later in the algorithm, the underlying data array is also changed. This is undesirable for two reasons: mutating the data array which is passed in could have downstream effects, and even if the data array isn't used subsequently, it makes the algorithm incorrect.

Assuming I'm not missing something here (very possible, as I am not a JS developer), data[idx].slice() would be the fix I would suggest.

Any chance to support esm?

I assume result.it should be <= maxIterations, but is rather a big number (close to MAX).

const skmeans = require("skmeans")
const data = [1,12,13,4,25,21,22,3,14,5,11,2,23,24,15];

const result = skmeans(data,3,null,10);
console.log(result.it) // 9991

I see that it is calculated using MAX - it, where MAX = 10000, and it was counted backwards from 10 in this case. Should rather use maxit if it is set.

@solzimer I was quite surprised that there was no predict method for the skmeans object. The clustering does a good job but if i want to assign new test examples to a cluster we would need a predict method.

I'm not sure if this has come up in the past but I've submitted some TypeScript definitions for the community at DefinitelyTyped/DefinitelyTyped#43059 if that's cool?

For me each run of the same dataset gives different results. Is this due to the kmpp initialisation algorithm? Is there a seed that can be passed in this case? Thanks.

I'm trying to import your library into a pure ES5 library, however your main in package.json is pointing to your ES6 main.js instead of your dist/skmeans.js (which is ES5 built by browserify/babel).

Would it possible to replace your main with your distributed file.

I'm using this library in a context where I need the results to be deterministic. Would it be possible to provide the option to pass a seed into the cluster initialization functions?

I was thinking an optional argument could be passed into the skmeans function called seed and that would be passed on to the kmpp and kmrand .

I am able to implement this functionality, I just want to make sure that this is behavior that library owners want.

Thanks!

Hi,

First off, love your package, it's helped me out a lot!

Secondly I was using your package to run some kmeans clustering and I noticed that the types published under the @types/skmeans package are not correct.

Specifically, the skmeans function is typed as returning Data and Data is typed as follows

interface Data {
    it: number;
    k: number;
    centroids: number;
    idxs: number[];
    test: (x: number, point?: (x1: number, x2: number) => number) => void;
}

https://github.com/adamzerella/DefinitelyTyped/blob/e84211cf43b463bb00759e3156ea7010516ee9b1/types/skmeans/index.d.ts#L11

But centroids: number is not correct. As far as I see the centroids property returned by skmeans returns an array of centroids, so the correct type should be centroids: number[] | number[][]