Clustering Refresher

When creating machine learning models, there can be two types of models. The ones previously discussed are with datasets that already have a target value to them. This means that multiple features together results in this outcome. However, clustering is part of the unsupervised models. This means that the model itself will not be told a target feature. When the dataset is put into the model, the model will give a result in terms of clustering or groups. These groups are ‘similar’ according to the parameters used and it is up to the creator to identify what the link between the data points within the group is.

Clustering can come in different forms or techniques used to group the data points. These most commonly include K-means and Hierarchical Agglomerative Clustering. When performing K-clustering, the model will need to decide on the amount of k-centers the model should be looking for. This means how many classes does the dataset need to end up in after running the model. The hard part of this model is usually defining what the k value is and should typically be decided before the model is run. Through out the process, the points will all attached to the nearest neighbor meaning a centroid point that can connect the data points together. The centroid point will keep moving until all data points are connected to one and the groups are identified. One of the methods to identify if the correct amount of k-centers was picked is to use the metric Calinski Harabaz Score, which is also known as the variance ratio. The score is calculated at multiple k values and the the graph will show an elbow shaped graph. The highest variance ratio is where the ideal k value is good to use for the model.

On the other hand, hierarchical agglomerative clustering is a different clustering method that uses the idea that all of the data points will start as their own cluster. So if you have 1000 data points, you start with 1000 clusters. There are multiple ways the data points are linked together in this model such as ward, average and complete. Ward takes notice of the lowest increase in variance between the clusters when combining. Average uses the smallest average distance between all the points in the cluster to make groups. Finally, complete uses the smallest maximum distance between the points between the points to make clusters. The resulting graph is that of a tournament bracket and when you draw a straight line across the bracket, the amount of lines that crosses on the bracket is the amount of clusters that the model is resulting in. A commonly used instance of HAC is the photos that are grouped together in smartphones today. They will all start on their own and then end up at a certain group for the phone to display.

In terms of clustering, a lot of instances can be used especially when the users do not have a concrete idea of what the similarities between all their data is and want to identify patterns.

Related posts:

Office wellness challenges are short-term behavior change contests, interventions, and team activities designed to improve health and employee morale. They can include: For example, one of our most…

Tennis is a beloved sport that’s great to play in the nice summer weather. It’s a fun game and although scoring sounds complicated, it’s pretty easy to learn the rules. One point is counted as 15…