Question

我正在尝试编写DBSCAN算法来聚类一组点，但我得到的结果非常糟糕。这可能是因为数据，但不仅如此。我正在获得大小的集群＆lt; minPoints不应该发生。

我做错了什么？我已多次查看代码，但我无法弄清问题是什么。

我参考了DBSCAN Wikipedia page上给出的算法。

private static int[] dbScan(String[] points, int epsilon, int minPts) {
    int cluster = 0;
    // visited stores if point has been visited
    boolean[] visited = new boolean[points.length];
    // pointsCluster stores which cluster a point has been assigned to
    int[] pointsCluster = new int[points.length];
    for(int iii = 0; iii < points.length; iii++) {
        // if point iii is already visited, do nothing  
        if(visited[iii]) continue;                      
        visited[iii] = true;    // mark point iii as visited
        // get points in neighborhood of point iii
        HashSet<Integer> neighbors = epsilonNeighbors(points, iii, epsilon);    
        if(neighbors.size() < minPts) {
            // if number of neighbors < minPts, mark point iii as noise
            pointsCluster[iii] = -1;
        } else {
            ++cluster;                      // else, start new cluster
            expandCluster(points, iii, neighbors, pointsCluster, visited, cluster, epsilon, minPts);
        }
    }
    return pointsCluster;
}

/*
 * Expands a cluster if a point is not a noise point
 * and has > minPts in its epsilon neighborhood
 */
private static void expandCluster(String[] points, int seedPoint, HashSet<Integer> neighbors,
        int[] pointsCluster, boolean[] visited, int cluster, int epsilon, int minPts) {

    pointsCluster[seedPoint] = cluster;     //assign cluster to seed point
    // create queue to process neighbors
    Queue<Integer> seeds = new LinkedList<Integer>();
    seeds.addAll(neighbors);
    while(!seeds.isEmpty()) {
        int currentPoint = (Integer) seeds.poll();
        if(!visited[currentPoint]) {
            visited[currentPoint] = true;       // mark neighbor as visited
            // get neighbors of this currentPoint
            HashSet<Integer> currentNeighbors = epsilonNeighbors(points, currentPoint, epsilon);
            // if currentPoint has >= minPts in neighborhood, add those points to the queue
            if(currentNeighbors.size() >= minPts) {
                seeds.addAll(currentNeighbors);
            }
        }
        // if currentPoint has not been assigned a cluster, assign it to the current cluster
        if(pointsCluster[currentPoint] == 0) pointsCluster[currentPoint] = cluster;
    }
}

/*
 * Returns a HashSet containing the indexes of points which are
 * in the epsilon neighborhood of the point at index == currentPoint
 */
private static HashSet<Integer> epsilonNeighbors(String[] points, int currentPoint, int epsilon) {
    HashSet<Integer> neighbors = new HashSet<Integer>();
    String protein = points[currentPoint];
    for(int iii = 0; iii < points.length; iii++) {
        int score = similarity(points[iii], points[jjj]);
        if(score >= epsilon) neighbors.add(iii);
    }
    return neighbors;
}

Answer 1

如果结果不好，可能是因为您的数据不好（基于密度的群集），或者因为您的参数不好。

事实上，DBSCAN 可以生成小于minPts的簇，如果它们相互接触的话。然后，他们可以“偷”彼此的边界点。

如何使用例如ELKI验证您的算法输出？

DBSCAN群集算法无法正常工作。我究竟做错了什么？

1 个答案: