【问题标题】：Mahout clustering - all text vectors in single cluster - why?Mahout 聚类 - 单个聚类中的所有文本向量 - 为什么？
【发布时间】：2019-01-20 20:32:24
【问题描述】：

我已经运行了以下示例：

https://github.com/technobium/mahout-clustering/blob/master/src/main/java/com/technobium/ClusteringDemo.java#L64

Document 1 -> John saw a red car.
Document 2 -> Marta found a red bike.
Document 3 -> Don need a blue coat.
Document 4 -> Mike bought a blue boat.
Document 5 -> Albert wants a blue dish.
Document 6 -> Lara likes blue glasses.
Document 7 -> Donna, do you have red apples?
Document 8 -> Sonia needs blue books.
Document 9 -> I like blue eyes.
Document 10 -> Arleen has a red carpet.

它可以与EuclideanDistanceMeasure 一起正常工作。但我不确定为什么文本预期的距离测量（TanimotoDistanceMeasure 和CosineDistanceMeasure）只给我一个集群。

这是为什么？我并没有假装我对这 2 个距离测量结果不满意的情况一无所知 - 但我可能需要改变什么？那里有太多数字，我无法理解每个数字的影响。我确实有《Mahout in Action》这本书，虽然我只读了两章。

EuclideanDistanceMeasure（2 个集群 - 好）

 Clusters: 
         7 -> wt: 1.0 distance: 4.4960791719810365  vec: Document 1 = [8:2.609, 21:2.609, 29:1.693, 30:2.609]
         7 -> wt: 1.0 distance: 4.496079376645949  vec: Document 10 = [2:2.609, 9:2.609, 18:2.609, 29:1.693]
         7 -> wt: 1.0 distance: 4.496079576525459  vec: Document 2 = [3:2.609, 16:2.609, 25:2.609, 29:1.693]
         9 -> wt: 1.0 distance: 4.389955960700927  vec: Document 3 = [4:1.357, 10:2.609, 13:2.609, 27:2.609]
         9 -> wt: 1.0 distance: 4.389956011306051  vec: Document 4 = [4:1.357, 5:2.609, 7:2.609, 26:2.609]
         9 -> wt: 1.0 distance: 4.3899560687101395  vec: Document 5 = [0:2.609, 4:1.357, 11:2.609, 32:2.609]
         9 -> wt: 1.0 distance: 4.389956137136399  vec: Document 6 = [4:1.357, 17:2.609, 22:2.609, 24:2.609]
         7 -> wt: 1.0 distance: 5.577549042707083  vec: Document 7 = [1:2.609, 12:2.609, 14:2.609, 19:2.609, 29:1.693, 33:2.609]
         9 -> wt: 1.0 distance: 4.389956708176695  vec: Document 8 = [4:1.357, 6:2.609, 28:2.609, 31:2.609]
         9 -> wt: 1.0 distance: 4.389471924190491  vec: Document 9 = [4:1.357, 15:2.609, 20:2.609, 23:2.609]

制作者：

    CanopyDriver.run(new Path(vectorsFolder), new Path(canopyCentroids), new EuclideanDistanceMeasure(), 20, 5,
            true, 0, true);

    FuzzyKMeansDriver.run(new Path(vectorsFolder), new Path(canopyCentroids, "clusters-0-final"),
            new Path(clusterOutput), 0.01, 20, 2, true, true, 0, false);

CosineDistanceMeasure（只有 1 个集群 - 不好）

Clusters: 
         0 -> wt: 1.0 distance: 0.6362357041216559  vec: Document 1 = [8:2.609, 21:2.609, 29:1.693, 30:2.609]
         0 -> wt: 1.0 distance: 0.6362357041216559  vec: Document 10 = [2:2.609, 9:2.609, 18:2.609, 29:1.693]
         0 -> wt: 1.0 distance: 0.636235704121656  vec: Document 2 = [3:2.609, 16:2.609, 25:2.609, 29:1.693]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 3 = [4:1.357, 10:2.609, 13:2.609, 27:2.609]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 4 = [4:1.357, 5:2.609, 7:2.609, 26:2.609]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 5 = [0:2.609, 4:1.357, 11:2.609, 32:2.609]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 6 = [4:1.357, 17:2.609, 22:2.609, 24:2.609]
         0 -> wt: 1.0 distance: 0.5876411474816594  vec: Document 7 = [1:2.609, 12:2.609, 14:2.609, 19:2.609, 29:1.693, 33:2.609]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 8 = [4:1.357, 6:2.609, 28:2.609, 31:2.609]
         0 -> wt: 1.0 distance: 0.6328896123664868  vec: Document 9 = [4:1.357, 15:2.609, 20:2.609, 23:2.609]

制作者

    CanopyDriver.run(new Path(vectorsFolder), new Path(canopyCentroids), new CosineDistanceMeasure(), 20, 5,
            true, 0, true);

    FuzzyKMeansDriver.run(new Path(vectorsFolder), new Path(canopyCentroids, "clusters-0-final"),
            new Path(clusterOutput), 0.01, 20, 2, true, true, 0, false);

TanimotoDistanceMeasure（只有 1 个集群 - 不好）

 Clusters: 
         0 -> wt: 1.0 distance: 0.8637279689324617  vec: Document 1 = [8:2.609, 21:2.609, 29:1.693, 30:2.609]
         0 -> wt: 1.0 distance: 0.8637279689324617  vec: Document 10 = [2:2.609, 9:2.609, 18:2.609, 29:1.693]
         0 -> wt: 1.0 distance: 0.8637279689324617  vec: Document 2 = [3:2.609, 16:2.609, 25:2.609, 29:1.693]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 3 = [4:1.357, 10:2.609, 13:2.609, 27:2.609]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 4 = [4:1.357, 5:2.609, 7:2.609, 26:2.609]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 5 = [0:2.609, 4:1.357, 11:2.609, 32:2.609]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 6 = [4:1.357, 17:2.609, 22:2.609, 24:2.609]
         0 -> wt: 1.0 distance: 0.8723755210900389  vec: Document 7 = [1:2.609, 12:2.609, 14:2.609, 19:2.609, 29:1.693, 33:2.609]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 8 = [4:1.357, 6:2.609, 28:2.609, 31:2.609]
         0 -> wt: 1.0 distance: 0.8596377086023765  vec: Document 9 = [4:1.357, 15:2.609, 20:2.609, 23:2.609]

通过

制作

    CanopyDriver.run(new Path(vectorsFolder), new Path(canopyCentroids), new TanimotoDistanceMeasure(), 20, 5,
            true, 0, true);

    FuzzyKMeansDriver.run(new Path(vectorsFolder), new Path(canopyCentroids, "clusters-0-final"),
            new Path(clusterOutput), 0.01, 20, 2, true, true, 0, false);

【问题讨论】：

在我看来，在那个玩具数据上，1 个集群是更好的结果。
我认为即使在我的真实数据上也会发生这种情况。有什么建议可以让第 11 个玩具文档获得第 2 个集群？
对于所有这些措施，您需要更长的文档才能使其正常工作。
哦。我正在尝试一些更大的文档（1-3 个非散文段落），但也只得到了一个集群。但是感谢您的反馈，我将使用数据集并尝试确定因果关系。
@Anony-Mousse - 非常感谢您的支持，事实证明您的第一条评论是对的。如果您将我的答案复制并粘贴为您自己的帖子，我会将其标记为已接受的答案，以便您获得信用。

标签： cluster-analysis mahout

【解决方案1】：

正如 Anony-Mousse 在他的第一个回复中所说，我提供给它的数据属于一个集群。在最近几周进行了一番深思熟虑（或者更具体地说，直接尝试了距离测量类）后，我发现了一个数据集，该数据集导致了多个集群：

1) 确保数据足够不同

Text id1 = new Text("Document 1");
Text text1 = new Text("Atletico Madrid win");
writer.append(id1, text1);

Text id6 = new Text("Document 6");
Text text6 = new Text("Both apple and orange are fruit");
writer.append(id6, text6);

Text id7 = new Text("Document 7");
Text text7 = new Text("Both orange and apple are fruit");
writer.append(id7, text7);

2) 确定好的半径值

a) 用您的样本数据试验 DistanceMeasure 类

Vector v1 = toVector("Atletico Madrid win");
Vector v2 = toVector("Both apple and orange are fruit");
Vector v3 = toVector("Both orange and apple are fruit");
of = ImmutableList.of(v1, v2, v3);

List<Vector> vectorList = new LinkedList();
vectorList.addAll(of);
List<Canopy> canopies = CanopyClusterer.createCanopies(vectorList, new CosineDistanceMeasure(), 0.3, 0.3);
for (Canopy canopy : canopies) {
    System.out.println("DistanceMeasureMain.main() " + canopy.asFormatString());
}

产生：

DistanceMeasureMain.main() distance is 0.19193857965451055
DistanceMeasureMain.main() distance is 0.5281191379648771
DistanceMeasureMain.main() distance is 0.19193857965451055
DistanceMeasureMain.main() C0: {0:1.1,117724:1.0,378550445:1.0,1997849123:1.0}
DistanceMeasureMain.main() C1: {0:1.1,96727:1.0,96852:1.0,2076577:1.0,93029210:1.0,97711124:1.0,1008851410:1.0}

b) 使用距离作为半径值

我认为CanopyDriver.run() 的t1 和t2 值（0.2 和0.2）也很重要，尽管我不知道调用中所有数字参数的影响的复杂细节下面：

    // CosineDistanceMeasure
    CanopyDriver.run(new Path(vectorsFolder),
            new Path(canopyCentroids), new CosineDistanceMeasure(),
            0.2, 0.2, true, 1, true);

    FuzzyKMeansDriver.run(new Path(vectorsFolder), new Path(
            canopyCentroids, "clusters-0-final"), new Path(
            clusterOutput), 0.01, 20, 2, true, true, 0, false);

输出

Document 1 -> Atletico Madrid win
Document 6 -> Both apple and orange are fruit
Document 7 -> Both orange and apple are fruit

 Clusters: 
         0 -> wt: 1.0 distance: 0.0  vec: Document 1 = [1:1.405, 4:1.405, 6:1.405]
         1 -> wt: 1.0 distance: 0.0  vec: Document 6 = [0:1.000, 2:1.000, 3:1.000, 5:1.000]
         1 -> wt: 1.0 distance: 0.0  vec: Document 7 = [0:1.000, 2:1.000, 3:1.000, 5:1.000]

【讨论】：