Comprehensive Guide to Cluster Analysis: Theory, Methods, and R Implementation

Introduction

The session introduced cluster analysis as an unsupervised technique for grouping similar observations into homogeneous clusters that are distinct from other groups. Participants were reminded to download the Day 5 materials (PowerPoint, PDF, and two R scripts – cluster_analysis.R and cluster_selfread.R) from the shared Google Drive.

Core Concepts

• Similarity vs. Dissimilarity: Within‑cluster distance should be small (high similarity); between‑cluster distance should be large (high dissimilarity).
• Homogeneity: Objects inside a cluster share common attributes (e.g., shape, color, material).
• Inter‑ and Intra‑cluster Distance: Intra‑cluster distance measures how close points are inside a cluster; inter‑cluster distance measures separation between clusters.
• Linkage Methods:
• Single linkage – uses the shortest distance between two clusters.
• Complete linkage – uses the longest distance.
• Average linkage – averages all pairwise distances.
• Visualization Tools:
• Dendrogram – tree‑like diagram from hierarchical clustering.
• Elbow (Scree) Plot – shows within‑cluster sum of squares to help select the optimal number of clusters.
• Scatter‑plot matrix – visual inspection of variable relationships.

Practical Example with Toy Data

A toy dataset (variables: shape, color, material) illustrated how choosing different attributes changes the number of clusters: - Grouping by shape yielded three clusters (triangle, circle, rectangle). - Adding color doubled the clusters because each shape split into red/blue groups. - Including material further increased cluster count. The example emphasized that the analyst decides which attributes to use; more attributes generally produce more clusters.

Hierarchical vs. Non‑hierarchical Clustering

• Hierarchical (Agglomerative): Starts with each observation as its own cluster and merges them step‑by‑step using a linkage function. The resulting dendrogram helps decide where to cut the tree.
• K‑means (Partitioning): Requires pre‑specifying k clusters. The algorithm iteratively updates centroids and reassigns points until convergence. Both methods were demonstrated in R.

Step‑by‑Step R Workflow

1. Load Packages – install and library required libraries (e.g., tidyverse, cluster).
2. Import Data – read the utilities.csv file from GitHub or a local copy.
3. Data Preparation –
4. Convert categorical columns (e.g., company) to factors.
5. Remove non‑numeric columns before scaling.
6. Normalization – apply scale() to obtain zero‑mean, unit‑variance variables, reducing noise from differing measurement units.
7. Distance Matrix – compute Euclidean distances with dist() on the normalized data.
8. Hierarchical Clustering – use hclust(dist_matrix, method = "complete") and plot the dendrogram.
9. Determine Optimal Clusters – inspect the elbow plot (fviz_nbclust) or cut the dendrogram at a chosen height.
10. K‑means Clustering – run kmeans(normalized_data, centers = k) for a predetermined k.
11. Interpret Results – examine cluster assignments, visualize with scatter plots, and discuss business implications (e.g., segmenting utility companies by fuel cost vs. sales).

Common Pitfalls & Troubleshooting

• File Access Errors – ensure internet connectivity or download the CSV locally and set the correct working directory.
• Package Installation Issues – run install.packages("packageName") before library().
• Incorrect Data Types – convert character columns to factors; exclude them from scaling.
• Choosing k – use domain knowledge, dendrogram inspection, or silhouette analysis to avoid arbitrary decisions.

Applications

Cluster analysis can be applied to market segmentation, genetic accession grouping, document clustering, image segmentation, and any scenario where natural groupings are sought without a predefined outcome variable.

Final Remarks

The session emphasized practice: participants should replicate the scripts, experiment with different variables, and explore alternative distance measures (Manhattan, Gower) to deepen understanding.

Cluster analysis empowers analysts to uncover natural groupings in data by balancing intra‑cluster similarity and inter‑cluster separation; mastering both hierarchical and K‑means methods in R, along with proper data preparation and validation, is essential for reliable segmentation across diverse fields.