1 2 3 4 5 6 7 8 9

Taxonomic attribute-values
For attributes which values are structured by relations of hierarchical type (classified values), an extension of the discrete classifier partitioning process is proposed.

Fig. 4. Classified values of attribute A

The method consists, when such a classifier is selected, in creating a set of partitions corresponding to the first level of the hierarchy (noted d_first = {v₁,..., v_i,..., v_k} with k elements). Each case is assigned to the partition that generalizes its value. Let A be a taxonomic attribute with the domain d = {v₁,..., v_i,..., v_n} of n modalities and  is a subtree of m submodalities of v_i [Fig. 4]:

Let Q be a Boolean application (called question) which determines if the modality v_i generalizes a value v_ij. Q is defined by:



Then, we can generate k partitions from d_first:



In the next step, we create temporarily k attributes {A₁,..., A_j,..., A_k} in each partition E_A1,..., E_Ak with a set of modalities defined by the subvalues of {v₁,..., v_i,..., v_k}. These ones can be picked by the test function (information gain, gain ratio) and the method is recursively reapplied.

Multi-valued attributes
When modeling the descriptive model, a discrete attribute (nominal or taxonomic) can be defined as multi-valued. It can express doubt (disjunction of imprecision) or the simultaneous presence of states (conjunction of variation) like in the following expression:

v = (v₁₁ & ... v_1i ... & v_1m) | ... | (v_j1 & ... v_ji ... & v_jn) | ... | (v_k1 & ... v_ki ... & v_kp) where cf_i = (v_j1 & ... v_ji ... & v_jn)

Depending on the semantic associated with a conjunctive form of a case (cf), IKBS can apply three processing methods:

1. If cf is true information (association of co-existing facts), create k partitions corresponding to each conjunction of v, and dispatch cases with such value in each partition: cf is seen as a new possible value of dom(A).
2. If cf expresses fuzzy information (the intrinsic variability of multiple objects is an adding source of noise), treat conjunctions as disjunction.
3. Allow the user to customize the degree of similarity  between two conjunctive forms.

The default method is the third one with  = 1 because it gives a good compromise between the tree size (number of nodes) and the discrimination accuracy. Indeed, the first method don't generate a deep tree, but carries a major risk of misidentification: each cf of the selected attribute at a node of the decision tree must match exactly the cf of the tested case. The third method is more flexible because it makes a fuzzy matching for dispatching cases in each partition, depending on the number of differences between the two conjunctive forms and .