Machine Learning: Clustering

Announcements

• Assignment 7: Due Wednesday May 18th
• Last class!

Clustering

Clustering

• Clustering is an unsupervised ML method
• Clustering algorithms will attempt to divide a set of data into 2 or more segments
• Data within clusters should be as similar as possible
• Data between clusters should be as different as possible

How can we best draw N boundaries around this data?

• Most algorithms require you to say how many clusters you want in advance
• But there are also methods to suggest an optimum number of clusters from the data

K-Means Clustering

• Simple, works fairly well
• Pick N random points in your data where N=number of clusters
• Each of these points is called the centroid
• Go through every other point and assign it to the nearest centroid
• For each of the clusters, find the new center of the cluster
• Now go through each point and re-assign to the closest centroid
• Keep doing this till the clusters stop changing

Uses of clustering

• Segmentation (audience/customer)
• Document clustering
• Image segmentation
• Recommendation engines

How "well" have we clustered data?

• Common measure: inertia
• To calculate:
  1. measure the distance between each data point and its centroid
  2. square this distance
  3. sum the squares across all points

Let's do some clustering...

DBSCAN

density-based spatial clustering of applications with noise

k-Means vs. DBSCAN

• density - number of points in a circle of specific radius (epsilon)
• minimum points - minimum density for a point to be considered a core point
• core point - points that meet minimum density metric
• border point - points that don't meet minimum density metric but are inside a circle for a core point
• noise - points that are not inside any other circle

Advantages over k-Means

• no need to a priori specify number of clusters in data
• can find clusters of arbitrary shapes
• has a concept of noise, good for data with outliers