IB is an instance-based inducer [,]. A good, robust algorithm, but still slow when there are many attributes.
NUM_NEIGHBORS determines the number of neighbors to use. In discrete domains, many neighbors will have the same distance, so NNKVALUE determines whether the number of neighbors will actually be neighbors of different distances, with tie-breaking instances counting as a single neighbor. NORMALIZATION determines whether the data should be normalized according to extreme values, according to the interquartile range (25% to 75%), or whether no normalization should take place. NEIGHBOR_VOTE determines whether the neighbors vote equally or with voting power inversely proportional to their distance. MANUAL_WEIGHTS allows setting the weights per attribute manually, i.e., by typing them for each attribute.