Clustering predictions using R language analysis


1. Overview

This article shows an example of using R for analysis, creating clusters using a K-means model. K-means clustering aims to partition a number (n) of observations into a number (k) of clusters, in which each observation belongs to the cluster with the nearest mean, serving as a prototype of the cluster. Clustering can help identify different groups in your data that should receive special treatment based on their characteristics (for example, a defined custom marketing campaign for a certain cluster).

This article is intended to show how you can use any R package or function for additional advanced analytics tasks. Dundas BI already includes a K-Means Clustering function you can use in metric set formulas, and can recommend the option to automatically identify clusters in a scatter plot.

2. Prepare the data

The following example uses the Iris data set. This data set contains sepal and petal measurements, which are used to predict the species of the iris plant by classifying the data points into different clusters.

If necessary, first install and configure R and Rserve.

3. Create the data cube using Rserve

Create a data cube that retrieves your data from any source, such as a database, Excel file, or a flat file such as the iris data file provided in CSV format.

The following shows the iris data in our example with column names assigned:

Iris data shown in Data Preview
Iris data shown in Data Preview

Using the Insert Other option in the toolbar, add an R Language Analysis transform after the Select transform.

Data Cube setup
Data Cube setup

Configure the R Language Analysis transform and add the following script:

#Define an argument for each column in the data set.
arg1 <- $PetalLength$
arg2 <- $PetalWidth$
arg3 <- $SepalLength$
arg4 <- $SepalWidth$
arg5 <- $Species$
arg6 <- $SampleID$

#Use the K-means algorithm to classify the Petal Length, Petal Width, Sepal Length, and Sepal Width into three clusters
result <- kmeans(data.frame(arg1, arg2, arg3, arg4), 3);

#Dundas BI requires that the output be in the form of a data frame
output <- data.frame(arg1, arg2, arg3, arg4, result$cluster, arg5, arg6);

This script is using the K-means clustering algorithm, which requires multiple arguments and defines the number of clusters to generate out of those variables. 

The number of clusters is an input parameter (in this example we used 3 as the number of clusters). An inappropriate choice of k may yield poor results. That is why, when using the K-means model, it is important to run diagnostic checks to determine the optimal number of clusters in the data set.

In the same configuration panel, click Define Element Placeholders and add the placeholders for all the arguments defined in the script. Make sure the column names match with something that the R script recognizes.

Define element placeholders
Define element placeholders

Open the data preview. You can see that the new result.cluster column, which is classifying each data point into clusters 1, 2, or 3.

Cluster result column added
Cluster result column added

arg5, the species we are trying to predict, isn't necessary for this example but is used here to help us assess the accuracy of the cluster detection by comparing it to known clusters - in this case the species of the iris flower.

4. Using SQL Server R Services

Alternatively, if you are using a database such as SQL Server or Azure SQL with R Services or Machine Learning Services, you may be able to run R script as part of the SQL query in a Manual Select transform.

Create a Manual Select data cube using the following script:

exec sp_execute_external_script  
      @language = N'R'    
    , @script = N'
		df <- data.frame(InputDataSet)
		arg1 <- df$SepalLength
		arg2 <- df$SepalWidth
		arg3 <- df$PetalLength
		arg4 <- df$PetalWidth
		arg5 <- df$Species
		arg6 <- df$SampleID

		result <- kmeans(data.frame(arg1, arg2, arg3, arg4), 3);
		OutputDataSet <- data.frame(arg1, arg2, arg3, arg4, result$cluster, arg5, arg6);
	, @input_data_1 = N'
Select 5.1 As [SepalLength], 3.5 As [SepalWidth], 1.4 As [PetalLength], 0.2 As [PetalWidth], ''I. setosa'' as [Species], 1 As [SampleID] UNION ALL
		Select 4.9, 3, 1.4, 0.2, ''I. setosa'', 2 UNION ALL
		Select 4.7, 3.2, 1.3, 0.2, ''I. setosa'', 3 UNION ALL
		Select 6.5, 3, 5.2, 2, ''I. virginica'', 148 UNION ALL
		Select 6.2, 3.4, 5.4, 2.3, ''I. virginica'', 149 UNION ALL
		Select 5.9, 3, 5.1, 1.8, ''I. virginica'', 150'    
WITH RESULT SETS (([Sepal Length] float, [Sepal width] float, [Petal length] float, [Petal width] float, [Cluster Result] int, [Species] varchar(MAX),[Sample ID] int));   

This script specifies the language (i.e., R) and executes the K-means algorithm as a standard stored procedure. It also includes the Select statements for retrieving the data points from the database.

The same result is obtained as with using Rserve:

Cluster result is the same as using R server
Cluster result is the same as using R server

5. Use the results

From the data cube created above, select the two measures, and add Sample ID to Rows in the Data Analysis Panel of a metric set.

Add the Predicted Cluster dimension to Columns, so that each Sample ID is grouped by the cluster.

Data Analysis Panel
Data Analysis Panel

Revisualize to a Scatter Plot. You will see that all the samples are grouped by cluster.

Scatter plot chart with clusters
Scatter plot chart with clusters

The clusters may change every time the data is retrieved (assuming that data is not cached). This may happen if the particular K-means algorithm used isn't deterministic and has a random component used to define the clusters.

6. See also

Dundas Data Visualization, Inc.
400-15 Gervais Drive
Toronto, ON, Canada
M3C 1Y8

North America: 1.800.463.1492
International: 1.416.467.5100

Dundas Support Hours:
Phone: 9am-6pm, ET, Mon-Fri
Email: 7am-6pm, ET, Mon-Fri