A Novel Feature Selection Strategy for Enhanced Biomedical Event Extraction Using the Turku System

Algorithm 1

AEE1 algorithm.

Step 1.  Categorize features into 𝑛 feature classes and there are 2^𝑛−1 possibilities for all kinds of feature

   combination. Denote 𝑁=2^𝑛−1, and run all of classifying test. 𝐹-score is recorded for

   each experiment.

Step 2. Sort 𝑁 feature combinations according to 𝐹-score with the descending order.

Step 3. FOR 𝑖=1 to 𝑛, 𝑗=1 to 𝑁,

      𝑂_𝑖𝑗=0, 𝑐_𝑖𝑗=0;

   END FOR

   // 𝑂_𝑖𝑗 is used to calculate the occurrence of the 𝑖th feature among 𝑗 experiments.

   // 𝑐_𝑖𝑗 evaluates the partial importance of the 𝑖th feature after 𝑗 times experiments.

Step 4. FOR 𝑗=1 to 𝑁, // For each feature combination in the 𝑗th experiment.

     FOR 𝑖=1 to 𝑛, // For each feature.

        IF the 𝑖th feature occur in the 𝑗th feature combination,

          𝑂_𝑖𝑗++;

          // For each experiment, the occurrence of the 𝑖th feature is accumulated.

        END IF

        𝑐_𝑖𝑗=𝑂_𝑖𝑗/𝑗;

        // By dividing 𝑂_𝑖𝑗 by 𝑗, we get a numerical value to evaluate the importance

        // of the 𝑖th feature. Since the items are ranked by 𝐹-score with a descending

        // order, the first ranked features combination corresponds to a smaller 𝑗 and

        // hence a bigger 𝑐_𝑖𝑗.

     END FOR

   END FOR

Step 5. FOR 𝑖=1 to 𝑛,

     AEE1(𝑖) =^𝑛∑_𝑗=1𝑐_𝑖𝑗;

   END FOR

   // By summing up 𝑐_𝑖𝑗 for a fixed 𝑖, we get the accumulated effect of the 𝑖th feature.

Step 6. AEE1(𝑖) is the objective function of greedy search algorithm, and we sort AEE1(𝑖)

   to get the most important feature.