Differences

This shows you the differences between two versions of the page.

--- sift:tutorials:run_k-means [2024/07/03 17:36] – created sgranger
+++ sift:tutorials:run_k-means [2024/11/05 14:58] (current) – wikisysop
@@ Line 1: / Line 1: @@
-====== Run_K-Means ======
+====== Run K-Means ======
-The k-means clustering algorithm is a commonly used method for grouping //n// individual data points into //k// clusters. It is a multi-variate statistical analysis that reduces the high-dimensional matrix of correlated, time-varying signals into a low-dimensional and statistically uncorrelated set of principal components (PCs). These PCs explain the variance found in the original signals and represent the most important features of the data, e.g., the overall magnitude or the shape of the time series at a particular point in the stride cycle. The value of each particular subject’s score for the individual PCs represents how strongly that feature was present in the data.
+The k-means clustering algorithm is a commonly used method for grouping //n// individual data points into //k// clusters. It does so in an unsupervised manner, iteratively selecting cluster centre points and assigning data points to a cluster. Within Sift, this is implemented onto the [[sift:application:analyse_page#workspace_scores|PC transformed data]], after a PCA analysis is done. More information about PCA can be found on our [[sift:principal_component_analysis:using_principal_component_analysis_in_biomechanics|Using PCA in Biomechanics page]].
 ==== The utility of clustering ====
-When analysing biomechanical signals, we often realize that a number of individual traces are similar. It can be useful to describe these traces as belonging to the same group, or cluster. This potentially allows us to simplify our analysis or to pick a single trace as being "representative" of the whole cluster. Because clustering is an unsupervised learning technique, it does not require any specific knowledge or set of training labels from the user. This, in turn, makes clustering useful for data exploration.
+When analysing biomechanical signals, we often realize that a number of individual traces are similar. It can be useful to describe these traces as belonging to the same group, or cluster. This potentially allows us to simplify our analysis or to pick a single trace as being "representative" of the whole cluster. Because k-means clustering is an unsupervised learning technique, it does not require any specific knowledge or set of training labels from the user. This, in turn, makes clustering useful for data exploration.
 ==== Tutorial Overview ====
-This tutorial works off the Principal Component Analysis Tutorial, and assumes a good understanding of using PCA in Sift. This tutorial uses overground walking data from roughly 100 subjects divided into two conditions, normal control and osteoarthritis (moderate to severe). This data set is included in the Demo folder of your Inspect3D installation (e.g., C:\Program Files\Sift\Demo). This is the same data as the PCA Tutorial.
+This tutorial works off the [[sift:tutorials:perform_principal_component_analysis|Principal Component Analysis Tutorial]], and assumes a good understanding of using PCA in Sift. This tutorial uses overground walking data from roughly 100 subjects divided into two conditions, normal control and osteoarthritis (moderate to severe). This data set is included in the Demo folder of your Sift installation (e.g., C:\Program Files\Sift\Demo). This is the same data as the PCA Tutorial.
 ==== Running a K-Means Test ====
-{{sift_new_kmeans.png}}
+{{:sift_new_kmeans.png}}
-  - Open the {{sift_outlier_detection.png}} **Outlier Detection Using PCA** dropdown on the toolbar.
+  - Open the {{:sift_outlier_detection.png}} **Outlier Detection Using PCA** dropdown on the toolbar.
   - Select K-Means in the dropdown.
   - Change the number of clusters to the correct number for your analysis. This can be an iterative approach, by conducting the K-means analysis multiple times until you are happy with the output. For this example we are going to stick to 2.
@@ Line 25: / Line 25: @@
 Once K-Means clustering is completed, in the dialog, you will be provided with a list of summary information on each cluster, including cluster center, cluster radius, and the workspaces that get grouped into each cluster.\\
-{{kmeansDlg.png}}
+{{:kmeansDlg.png}}
 ==== Viewing K-Means Results ====
-{{DataDlg.png}}
+{{:k-means-datadlg.png}}
 Once you have run your K-Means Test and taken a brief look at the cluster's summary statistics, you want to be able to visualize your clusters.
-  - Open up the {{sift_data_options.png}} **Data Options** dialog.
+  - Open up the {{:sift_data_options.png}} **Data Options** dialog.
   - In the top right corner under **Display Styles From...** make sure that **Cluster** is selected.
   - From the **Data Options** dialog you are also able to change the color or style of each cluster. Select **Clusters** in the **Edit Styles From** list on the left and play around with editing the styles of each cluster.
   - Navigate to the **Analyse** page and select the **Workspace Scores** tab in your PCA results.
-Looking at the workspace tab we can select different points and the group and file will be displayed. This allows us to view which data points in a cluster belong to what group. We can clearly see the data points split into to clusters, blue and red, with somewhat of a separation. red tends to be in the top and blue tends to be in the bottom.
+Looking at the workspace tab we can select different points and the group and file will be displayed. This allows us to view which data points in a cluster belong to what group. We can clearly see the data points split into to clusters, blue and red, with somewhat of a separation. Red tends to be in the top and blue tends to be in the bottom.
-{{WorkspaceScores.png}}
+{{:k-means-workspacescores.png}}
 A K-means test finds the similarity between data points and groups them together into clusters. If you had two groups that were vastly different, the clusters would not have mixed groups. If the data points between groups have similarities the clusters may have data points from different groups.
@@ Line 46: / Line 46: @@
 If we look at the results from this K-Means we can see that the clusters are not a perfect representation of each group, signifying that there is some overlap and similarities between groups. The graph on the left shows the groups split up with the osteoarthritis group in purple and the normal group in green. The graph on the right shows the two clusters. If we look at the red cluster it seems to be mainly the osteoarthritis group, and if we look at the blue cluster it seems to be mainly the normal group. The points circled in red show some osteoarthritis datapoints in the second cluster, again indicating some overlap.
-{{sift_groups_clusters.png}}
+{{:sift_groups_clusters.png}}
 ==== Reference ====