Method: Nearest neighbours

There are two important questions:

• How many neighbours should we choose?
  • Less neighbours \(=\) more local is the analysis.
  • More neighbours \(=\) less variability of the results.
  • Having about 20 neighbours could work fine (depending on dataset size).
• What metric should be used to measure the “proximity” of observations?
  • If the data is numeric we could use:
    • Euclidean distance
    • Manhattan distance
    • Minkowski distance
    • Chebyshev distance
    • Cosine similarity
  • As we have more predictors then the results will change from metric to metric.