Measures of Clustering Quality

Two Questions in Machine Learning

A bibliography can do more than list papers. It can show which questions belong to a topic. In this section, the two references point to two different machine learning questions: one asks how to evaluate clustering algorithms, while the other asks about the learnability of bipartite ranking functions.

These references are similar because each identifies a focused research question. They differ in subject: the clustering reference concerns the quality of grouped data, whereas the ranking reference concerns whether bipartite ranking functions can be learned effectively.

Assignment Versus Evaluation

Clustering groups similar data points into clusters. That grouping is the clustering process itself: data points are assigned to groups. Measures of clustering quality address a different stage. They are used to evaluate clustering algorithms and to judge the quality of the resulting clustering. Keeping these stages separate prevents a common mistake: treating the act of forming clusters as if it were already evidence that the clusters are good.

Tracing a Clustering Reference

The important reading move is to distinguish what the reference evaluates from the data-grouping process itself. The clustering reference is not being classified merely because it mentions clusters; it is being classified more precisely because it focuses on how clustering algorithms can be judged.

Learnability in Ranking

Ranking functions are used to rank data. A bipartite ranking function is a specific focus of the ranking reference. That reference asks a different question from the clustering reference: not how to evaluate a clustering algorithm, but how well bipartite ranking functions can be learned. Learnability therefore shifts attention from the output ordering alone to the possibility of learning the ranking function effectively.

Where the Topics Meet

Clustering quality measures and ranking functions are connected because both concern the analysis of machine learning results or functions. They differ in what they evaluate or produce. Clustering groups similar data points into clusters, and clustering quality measures evaluate clustering algorithms. Ranking functions rank data, while the ranking reference focuses on whether bipartite ranking functions can be learned.

Common Classification Mistakes

• Treating clustering quality as another name for clustering.

  Grouping data describes clustering, while evaluating clustering algorithms describes the role of quality measures.

  Fix: Describe clustering as the grouping process and clustering quality as the evaluation of that process or its output.

• Classifying both references as clustering references because they appear together.

  The references address different subjects: clustering quality and learnability of bipartite ranking functions.

  Fix: Read each reference for its own object of study and stage of analysis.

• Interpreting ranking as the same task as clustering.

  The source distinguishes ranking functions, which rank data, from clustering, which groups similar data points into clusters.

  Fix: Use ranking for ordering data and clustering for grouping similar data points.

• Ignoring the word learnability in the ranking reference.

  Its specific focus is how well bipartite ranking functions can be learned.

  Fix: Classify it more precisely under learnability of bipartite ranking functions.

Practice the Distinction

Key Takeaways

1. Clustering groups similar data points into clusters.
2. Measures of clustering quality evaluate clustering algorithms rather than merely performing the grouping.
3. The Ackerman and Ben-David reference belongs primarily to clustering quality and evaluation.
4. The ranking reference belongs primarily to the learnability of bipartite ranking functions.
5. The two references are related as focused machine learning questions, but they address different subjects and different stages of analysis.