Understanding Unsupervised Learning

Unsupervised Learning

• Unsupervised learning does not provide labeled examples of correct answers.
• It allows the learning algorithm to present a structure for human review.
• This is a highly iterative process to find meaningful patterns and relationships.

Measuring Unsupervised Learning

• There are no metrics to measure unsupervised learning.
• Questions to consider include whether there is an informative way to visualize the data and whether subgroups can be discovered among the variables or observations.

Human Supervision in Unsupervised Learning

• Unsupervised learning requires human supervision.
• The algorithm will perform the search, but humans select the learning algorithm, distance metrics, and feature selection.
• Unsupervised learning can be used as part of exploratory data analysis (EDA).

Clustering and Segmentation

• Clustering is also known as a segmentation technique, which involves dividing data into separate parts.
• There is no single correct answer in clustering.
• The approach depends on the goals and is constrained by the available data.
• Clustering is based on similarity, with the goal of identifying homogeneous subsets based on similarity within the subset and the number of subsets.

Similarity

• For numeric variables, similarity is based on distance (the delta between values).
• For categorical variables, similarity is based on having the same values.
• Each feature (column) in a dataset represents a dimension with a potential for similarity between instances (rows).
• Distance between features is measured to find "density" or points of concentration that form clusters.

Boxplots and Outliers

• Distance-based algorithms are sensitive to outliers, which are points beyond the "whiskers" in a boxplot.

Feature Scaling

• Outliers are typically addressed through feature scaling.
• Feature scaling transforms values to a common scale, such as a range of 0 to 1.
• Feature scaling ensures all features have an opportunity to contribute to the model, preventing larger-scale features from dominating.

Cluster Analysis Methods

• Two common methods for cluster analysis are hierarchical clustering and K-Means.

Hierarchical Clustering

• Hierarchical clustering is an iterative process.
• It starts with one cluster, splits it, and continues splitting until complete.
• Hierarchical clustering is computationally intensive and difficult on large datasets.
• It is best used when meaningful to a small dataset, such as looking at stores versus customers.

K-Means Clustering

• K-Means clustering is used for numerical data.
• Data usually consists of a set of measurements about objects of interest.
• The input must be numerical, with a defined distance metric, such as Euclidean distance.
• The output includes the centers of each discovered cluster and the assignment of each input datum to a cluster, known as a centroid.

K-Means Measures of Distance

• K-means defines two measures of distances: the distance between two data points (records) and the distance between two clusters.
• Distance can be calculated in various ways, but four principles tend to hold true:
  • Distance is not negative
  • Distance from one record to itself is zero
  • The distance from record I to record J is the same as from record J to record I
  • The distance between two records cannot be greater than the sum of the distances between each record and a third record

Unsupervised Learning Use Cases

• Unsupervised learning is often used as an exploratory technique to discover structure in the data.
• The technique can also summarize the properties of each cluster.
• Sometimes used as a prelude to classification to discover classes.
• Examples include household income, yearly purchase amount in dollars, and the number of household members of customer households.