Bio-medical Ontologies Maintenance and Change Management

Extraction of Constraints from Biological Data 181
called itemset. We are interested in finding frequent itemsets, which are itemsets
with support higher than a specific threshold.
A well-known technique for frequent itemset extraction is the Apriori algorithm
[1]. It relies upon a fundamental property of frequent itemsets, called the apriori
property: every subset of a frequent itemset must also be a frequent itemset.
The algorithm proceeds iteratively, firstly identifying frequent itemsets containing
a single item. In subsequent iterations, frequent itemsets with n items identified in
the previous iteration are combined together to obtain itemset with n+1 items. A
single scan of the database after each iteration suffices to determine which generated
candidates are frequent itemsets.
Once the largest frequent itemsets are identified, each of them can be subdivided
into smaller itemsets to find association rules. For every largest frequent
itemset s, all non-empty subsets a are computed. For every such subset, a rule
a →(s-a) is generated and its confidence is computed. If the rule confidence exceeds
a specified minimum threshold, the rule is included in the result set.
5.3 Quasi Tuple Constraints
For tuple constraints, the minimum confidence must be equal to 1. The minimum
support value allows us to concentrate on the most frequent constraints. If the support
is set to the inverse of the total number of records, then all the constraints are
considered (this support corresponds to rules contained in a single data entry). We
defined the tolerance as the complement of the confidence value (e.g., if confidence
is 0.95, then tolerance is 0.05).
By setting a tolerance value of 0.0, our method detected all and only the tuple
constraints contained in the SCOP and in the CATH databases. Furthermore, we
performed experiments with tolerance values ranging from 0.001 to 0.100 (corresponding
to confidence values from 0.999 to 0.900), while the support value has
been set to the inverse of the total number of records. Results show a consistent
number of quasi tuple constraints whose presence increases as the tolerance value
increases (see Fig. 3).
By investigating tuples that violate such constraints, anomalies can be identified.
Such analysis allows experts interested in the domain to focus their study on
a small set of data in order to highlight biological exceptions or inconsistencies in
the data. As the results in Fig. 3 show, the lower the tolerance is, the stronger the
tuple constraint between the two values is and the fewer rules are extracted. Tuning
the tolerance, a domain expert is allowed to concentrate on the desired subset
of most probable anomalies present in the data. As an example, on the SCOP database
a tolerance value of 0.05 extracts less than 200 rules to be further investigated
by domain experts, among the overall 30000 protein entries.
To distinguish between biological exceptions or inconsistencies, a further
analysis of the discovered anomalies can be performed by means of three
approaches: