✳flâneur — a map of the web's best reading

Classification

fairmlbook.org · 13,349 words · saved by 1 readers

The goal of classification is to leverage patterns in natural and social processes to conjecture about uncertain outcomes. An outcome may be uncertain because it lies in the future. This is the case when we try to predict whether a loan applicant will pay back a loan by looking at various characteristics such as credit history and income. Classification also applies to situations where the outcome has already occurred, but we are unsure about it. For example, we might try to classify whether financial fraud has occurred by looking at financial transactions. What makes classification possible is the existence of patterns that connect the outcome of interest in a population to pieces of information that we can observe. Classification is specific to a population and the patterns prevalent in the population. Risky loan applicants might have a track record of high credit utilization. Financial fraud often coincides with irregularities in the distribution of digits in financial statements. T

Home Modeling populations as probability distributions Formalizing classification Optimal classification Risk scores Varying thresholds and ROC curves Supervised learning Groups in the population No fairness through unawareness Statistical non-discrimination criteria Independence Limitations of independence Separation Why equalize error rates? Visualizing separation Conditional acceptance rates Sufficiency Calibration and sufficiency Calibration by group as a consequence of unconstrained learning How to satisfy a non-discrimination criterion Relationships between criteria Independence versus s

Explore this link on the map →

saved by

related reading