Clustering

• In search of minority (low probability) classes

~ 100 objects per 10⁶ - 10⁷ sky objects

• E.g. high-redshift quasars (z>4)
• Using SKICAT discovered 20 new z>4 quasars
• reduced observation time by factor of 40.
• How does one find minority classes?
• Most clustering algorithms would ignore them as noise or undesirable side-effects
• sampling is NOT useful
• Need to scale to > millions of cases

---