Case selection and interpolation in CBR retrieval

In this paper, several existing interpolation methods of use in CBR are discussed and compared. Interpolation in CBR is normally applied to a retrieval set of cases which are ‘near to’ a given target set in the problem domain. The interpolation method is then used to select an appropriate solution value from a solution domain. The main factors examined here, governing the accuracy and power of the interpolation, are the selection of cases for interpolation and the method of interpolation. Two selection criteria are examined: selection by nearest neighbours and selection by divergence algorithms. Three interpolation methods examined are examined: nearest neighbour, distance weighted nearest neighbour, linear regression and a generalised regression method, suitable to nominal values. Experimental results on three case-bases are presented for comparison. These are a: a real valued 2- dimensional sinusoidal random valued function, the classical iris case base, and the travel case base. The results show that linear regression is best for dense case bases, but is limited to real continuous problems. For general CBR usage, divergence selection can improve accuracy by a factor of 2, and that generalised regression can additionally improve accuracy also by a factor of 2.