K-Medoids (PAM) Clustering

K-medoids clustering, commonly fitted with the PAM (Partitioning Around Medoids) algorithm, partitions observations into k clusters each represented by an actual data point that minimizes total dissimilarity to its members.

K-메도이드 군집분석은 흔히 PAM 알고리즘으로 적합하며, 관측값을 k개 군집으로 나누되 각 군집을 구성원과의 총 비유사도가 최소인 실제 관측값으로 대표한다.

How do I cluster observations in a way that is robust to outliers or works with any distance measure?

이상치에 강하거나 어떤 거리척도로도 쓸 수 있는 군집분석은 어떻게 하나요?

K-medoids clustering with the PAM algorithm groups observations around representative data points rather than averages, which makes it less sensitive to outliers than k-means. It also accepts any dissimilarity, such as Gower distance for mixed numeric and categorical variables.

When to use K-Medoids (PAM) Clustering

• Grouping observations on several numeric or mixed-type features
• Data with outliers that would distort k-means centroids
• Needing real, interpretable cluster representatives

When to use something else

• Very large datasets where PAM is slow; use CLARA or k-means
• Clusters of irregular shape or unknown number; use DBSCAN or hierarchical clustering

Assumptions and requirements

• Number of clusters k chosen in advance
• Meaningful dissimilarity measure
• Scaling of features when distances combine different units

What the analysis reports

• Cluster assignments
• Medoid observations
• Average silhouette width
• Total within-cluster dissimilarity
• Cluster profile table

How to report it

K-medoids clustering (PAM) yielded [k] clusters with an average silhouette width of [value].

Related methods

• K-Means Clustering
• Hierarchical Clustering
• CLARA

Run K-Medoids (PAM) Clustering on Quorant