Classifier. The classifier tries to identify the query-topic using query-terms and training data consisting of topic- descriptions discussed above. In our implementation we use a multinomial NBC, with maximum likelihood estimates to determine the topic probabilities of the query. For a query q, we compute the probability of membership of the query for different topic-classes as, P (c |q) = P (q|ci) × P (ci) ∝ P (c ) Y P (q |c ) (6) where TSRki is the topic-sensitive SourceRank score of source sk for topic-class ci. CSRs give the query- topic sensitive SourceRank for all deep-web sources. Since CSR is computed during query-time, it is im- portant that its processing time is kept to a minimal. CSR will be used in conjunction with a relevance mea- sure as described below. Hence CSR computation can be limited to selected top-k most relevant topics.
Classifier. To determine which cluster is most suggestive for malig- xxxxx, we use an iterative training procedure. This method was first described by Xxxxxxxxxxxx et al (12). The following steps are performed until a repetitive sequence is reached: (1) Train an initial classifier C0 from all clusters from all lesions, using linear discrimi- nant analysis (LDA) and cross validation. This classifier will perform sub-optimally. (2) Using C0, find for each lesion the cluster with the highest probability of malig- xxxxx. From this collection of one cluster per lesion, train a new classifier C1 using a support vector machine and cross validation. (3) Iteratively determine classifiers Cj in a similar way as step 2. Stop when Cj + 1 [ C1.. .j, i.e., a repetitive sequence of classifiers is reached.
