为什么在切换循环时得到NaN？

Question

I'm trying to write a k-means function in javascript. And here is my code.

function kmeans(arrayToProcess,cluster_n){
    var pointDimension = arrayToProcess[0].length;
    var ClusterResult = new Array();
    var ClusterCenter = new Array();
    var oldClusterCenter = new Array();
    var changed=false;
    for(var i = 0;i<cluster_n;i++)
        ClusterCenter.push(arrayToProcess[randomInt(arrayToProcess.length-1)]);

    console.log(ClusterCenter);

    // do{
    for(var k=0;k<50;k++){//loop
        for(var i = 0; i<cluster_n; i++){
            ClusterResult[i] = new Array();
        }
        for(var i = 0; i<arrayToProcess.length; i++){
            //for every point element
            var oldDistance=-1;
            var newClusterNumber = 0;
            for(var j = 0; j<cluster_n; j++){
                //for every cluster
                var distance = Math.abs(computeDistanceBetween(arrayToProcess[i], ClusterCenter[j]));   
                if (oldDistance == -1){
                    oldDistance = distance;
                    newClusterNumber = j;
                }else if ( distance <= oldDistance ){
                    newClusterNumber = j;
                    oldDistance = distance;
                }
            }
            ClusterResult[newClusterNumber].push(arrayToProcess[i]);
        }
        oldClusterCenter = ClusterCenter;
        //compute new centroid
        for(var i = 0; i<cluster_n; i++){
            newCentroid = pinit(pointDimension);
            for(var j = 0; j<ClusterResult[i].length; j++){
                newCentroid = padd(ClusterResult[i][j], newCentroid);
            }
            ClusterCenter[i] = pdivide(newCentroid, ClusterResult[i].length);
        }

        changed=false;
        for(var i = 0; i<cluster_n; i++){
            if(!pequal(ClusterCenter[i],oldClusterCenter[i]))
                changed = true;
        }
    }//while (changed == true);

    return ClusterResult;
}


function computeDistanceBetween(a,b){
    var result = 0;
    for(var i = 0; i<a.length;i++) result += a[i] * b[i];
    return result;
}

function pinit(n){
    var result = new Array(n);
    for(var i=0;i<n;i++) result[i] = 0;
    return result;
}

function padd(a,b){
    var result = new Array(a.length);
    for(var i = 0; i<a.length;i++) result[i] = a[i] + b[i];
    return result;
}

function pdivide(a,d){
    var result = new Array(a.length);
    for(var i = 0; i<a.length;i++) result[i] = a[i] / d;
    return result;
}

function pequal(a,b){
    for(var i = 0; i<a.length;i++) 
        if(a[i] != b[i]) return false;
    return true;
}

function randomInt(max){
    return randomIntBetween(0,max);
}

function randomIntBetween(min,max){
    return Math.floor(Math.random() * (max - min + 1)) + min;
}

If I stop the for-loop(k<0), the console gives the right answer. But if I start the for-loop(k<1),the array ClusterCenter will always has some NaN items. How dose the NaN appear?

Edit: Further explanation: if the for-loop in the 14th line has been executed, the ClusterCenter above will give some NaN items.Why?

Example input

var testArray = new Array();
for(var i=0; i<100; i++) testArray.push([randomInt(-150,150),randomInt(-150,150)]);
kmeans(testArray,4);

Answer 1

the ClusterCenter above will give some NaN items.Why?

Because you're diving zero by zero, which is not a number. This does happen for every empty cluster in the ClusterResult - it will create ClusterCenter[i] = pdivide(pinit(pointDimension), 0); .

How to deal with empty clusters? Possible strategies I could think of would be to make 0/0 = 0 , to choose a new random cluster center, or to drop the cluster alltogether ( cluster_n-- ).

But why do you get so many empty clusters in the first place? Because your computeDistanceBetween function is seriously flawed. Every (non-0|0) point is distant from itself . Choose a more reasonable distance function, like euclidian distance. It should always return a positive number, rendering the Math.abs in the loop superflouos.

---

Some other points:

• newCentroid misses a var statement and leaks into global scope

• Your changed is flawed. When setting oldClusterCenter = ClusterCenter , both variables will hold the same array that is then mutated. Not only is pequal(ClusterCenter[i],oldClusterCenter[i]) always true, but even ClusterCenter[i]===oldClusterCenter[i] because of oldClusterCenter === ClusterCenter .

  To fix this, either make oldClusterCenter = ClusterCenter.slice() or introduce ClusterCenter = new Array(cluster_n); after the assignment.

• Your code for computing the nearest cluster could be simplified to

   var newClusterNumber = 0, oldDistance = computeDistanceBetween(arrayToProcess[i], ClusterCenter[0])); for (var j=1; j<cluster_n; j++) { var distance = computeDistanceBetween(arrayToProcess[i], ClusterCenter[j]); if (distance <= oldDistance) { newClusterNumber = j; oldDistance = distance; } } 

  or

   var onewClusterNumber, ldDistance=Infinity; for (var j=0; j<cluster_n; j++) { var distance = computeDistanceBetween(arrayToProcess[i], ClusterCenter[j]); if (distance <= oldDistance) { newClusterNumber = j; oldDistance = distance; } }