Scoring metrics reference

.ml.clust Scoring metrics

Unsupervised learning daviesbouldin Davies-Bouldin index dunn Dunn index silhouette Silhouette score

Supervised learning homogeneity Homogeneity score between predictions and actual value

Optimum number of clusters elbow Distortion scores for increasing numbers of clusters

KxSystems/ml/clust/score.q

Scoring metrics allow you to validate the performance of your clustering algorithms in three distinct use cases.

use case	analysis
Unsupervised learning	These metrics analyze how well data has been assigned to clusters, measuring intra-cluster similarity (cohesion) and differences (separation). In general, clustering is said to be successful if clusters are well spaced and densely packed. Used when the true cluster assignment is not known.
Supervised learning	If the true and predicted labels of the dataset are known, clusters can be analysed in a supervised manner by comparing true and predicted labels.
Optimum number of clusters	The optimum number of clusters can be found manually in a number of ways using techniques above. If the required number of clusters is not known prior to clustering, the Elbow Method is used to estimate the optimum number of clusters within the dataset using K-means clustering.

.ml.clust.daviesbouldin

Davies-Bouldin index

.ml.clust.daviesbouldin[data;clt]

Where

• data represents the points being analyzed in matrix format, where each column is an individual datapoint
• clt is the list of clusters returned by one of the clustering algorithms in .ml.clust

returns the Davies-Bouldin index, where a lower value indicates better clustering, with well-separated, tightly-packed clusters.

q)show d:2 10#20?10.
4.126605 8.429965 6.214154 5.365242 7.470449 6.168275 6.876426 6.123797 9.363..
4.45644  7.274244 1.301704 2.018829 1.451855 9.819545 7.490215 6.372719 5.856..
q)show r1:10?3
0 1 2 0 1 0 0 1 0 1
q)show r2:10?3
2 2 1 0 2 2 1 2 0 0

q)// lower values indicate better clustering
q).ml.clust.daviesbouldin[d;r1]
9.014795
q).ml.clust.daviesbouldin[d;r2]
5.890376

The Davies-Bouldin index works by calculating the ratio of how scattered data points are within a cluster, to the separation that exists between clusters.

.ml.clust.dunn

.ml.clust.dunn[data;df;clt]

Where

• data represents the points being analyzed in matrix format, where each column is an individual datapoint
• df is the distance function as a symbol, e.g. e2dist edist mdist
• clt is the list of clusters returned by the clustering algorithms in .ml.clust

returns the Dunn index, where a higher value indicates better clustering, with well-separated, tightly-packed clusters.

q)show d:2 10#20?10.
3.927524 5.170911 5.159796  4.066642 1.780839 3.017723 7.85033  5.347096..
4.931835 5.785203 0.8388858 1.959907 3.75638  6.137452 5.294808 6.916099..

q)show r1:10?3
0 0 1 1 0 0 2 0 1 0
q)show r2:10?3
0 0 1 1 0 2 0 2 1 2

q)// higher values indicate better clustering
q).ml.clust.dunn[d;`edist;r1]
0.5716933
q).ml.clust.dunn[d;`e2dist;r2]
0.03341283

The Dunn index is calculated based on the minimum inter-cluster distance divided by the maximum size of a cluster.

.ml.clust.elbow

The elbow method: a distortion score for each value of k applied to data, using k-means clustering.

.ml.clust.elbow[data;df;k]

Where

• data represents the points being analyzed in matrix format, where each column is an individual datapoint
• df is the distance function as a symbol, e.g. e2dist edist
• k is the maximum number of clusters

returns distortion scores for each set of clusters produced by k-means, with increasing values of k up to the user defined value.

q)show d:2 10#20?10.
3.927524 5.170911 5.159796  4.066642 1.780839 3.017723 7.85033  5.347096..
4.931835 5.785203 0.8388858 1.959907 3.75638  6.137452 5.294808 6.916099.. 
q).ml.clust.elbow[d;`edist;5]
16.74988 13.01954 10.91546 9.271871

If the values produced by .ml.clust.elbow are plotted, it is possible to determine the optimum number of clusters. The above example produces the following graph

It is clear that the elbow score occurs when the data should be grouped into 3 clusters.

.ml.clust.homogeneity

Homogeneity score

.ml.clust.homogeneity[pred;true]

Where

• pred is the predicted cluster labels
• true is the true cluster labels

returns the homogeneity score, bounded between 0 and 1, with a high value indicating a more accurate labeling of clusters.

q)show true:10?3
2 1 0 0 0 0 2 0 1 2
q)show pred:10?3
2 1 2 0 1 0 1 2 0 1
q).ml.clust.homogeneity[pred;true]
0.225179
q).ml.clust.homogeneity[true;true]
1f

Homogeneity score works on the basis that a cluster should contain only samples belonging to a single class.

.ml.clust.silhouette

Silhouette coefficient

.ml.clust.silhouette[data;df;clt;isavg]

Where

• data represents the points being analyzed in matrix format, where each column is an individual datapoint
• df is the distance function as a symbol, e.g. e2dist edist mdist
• clt is the list of clusters returned by the clustering algorithms in .ml.clust
• isavg is a boolean - 1b to return the average coefficient, 0b to return a list of coefficients

returns the Silhouette coefficient, ranging from -1 (overlapping clusters) to +1 (separated clusters).

q)show d:2 10#20?10.
3.927524 5.170911 5.159796  4.066642 1.780839 3.017723 7.85033  5.347096..
4.931835 5.785203 0.8388858 1.959907 3.75638  6.137452 5.294808 6.916099..

q)show r1:10?3
0 0 1 1 0 0 2 0 1 0
q)show r2:10?3
0 0 1 1 0 2 0 2 1 2

q)// Return the averaged coefficients across all points
q).ml.clust.silhouette[d;`edist;r1;1b]
0.3698386
q).ml.clust.silhouette[d;`e2dist;r2;1b]
0.2409856

q)// Return the individual coefficients for each point
q).ml.clust.silhouette[d;`e2dist;r2;0b]
-0.4862092 -0.6652588 0.8131323 0.595948 -0.2540023 0.5901292 -0.2027718 0.61..

The Silhouette coefficient measures how similar an object is to the members of its own cluster when compared to other clusters.