Find natural groupings in data using multiple clustering approaches, evaluate quality, and interpret results.
-
Explore features. Call describe_database with the database name from $ARGUMENTS. Identify numeric columns suitable for clustering. Note any categorical columns that could provide context for interpreting clusters later.
-
Extract and review data. Call execute_query to select the numeric feature columns. Check for nulls and extreme outliers in the sample. Note the number of observations and features.
-
Run K-Means clustering. Call analyze_clusters with the database name, feature columns, and algorithm set to "kmeans". Start with k=3 unless domain knowledge suggests otherwise. Review the silhouette score and cluster sizes.
-
Try alternative k values. If the silhouette score is below 0.5, re-run analyze_clusters with k=2, k=4, and k=5. Compare silhouette scores to find the optimal number of clusters.
-
Run DBSCAN for comparison. Call analyze_clusters with algorithm set to "dbscan". This density-based approach does not require specifying k and can find irregularly shaped clusters. Compare the number of clusters found and the noise point percentage.
-
Evaluate and compare. Assess both approaches:
- Silhouette scores (higher is better, above 0.5 is good)
- Cluster balance (are clusters roughly even or heavily skewed?)
- Number of noise points in DBSCAN
- Which method produces more interpretable groupings
-
Reduce dimensions for visualization. Call reduce_dimensions with the database name and feature columns, using PCA with 2 components. This provides a 2D representation of the clusters for interpretation.
-
Interpret clusters. For the best clustering result, describe each cluster by its feature averages. Give each cluster a descriptive label based on its defining characteristics. Note which features most differentiate the clusters.
-
Present results. Provide:
- Recommended number of clusters and algorithm
- Cluster profiles with feature summaries
- Silhouette score and quality assessment
- Observations about cluster separation and overlap
-
Recommend next steps. Suggest using cluster labels as a feature in /localdata-mcp:regression, running /localdata-mcp:analyze-correlations within each cluster, or applying the clustering to new data for segmentation.
