Gene Tree Labeling Using Nonnegative Matrix Factorization on Biomedical Literature

Algorithm 1

Feature vector replacement algorithm.

Input: MeSH Term-by-Document Matrix 𝐴^′_{𝑚′×𝑛}

    Factor Matrices 𝑊_𝑚×𝑘 and 𝐻_𝑘×𝑛 of original Term-by-Document Matrix 𝐴_𝑚×𝑛

    Global weight vector 𝑔^′,

    Threshold 𝑟 number of MeSH headings to represent each document

Output: MeSH feature matrix 𝑊^′

for 𝑖=1∶𝑛do

      Choose 𝑟 top globally-weighted MeSH headings from 𝑖th column of 𝐴^′

      Determine 𝑗=argmax_𝑗<𝑘𝐻_𝑗𝑖

      for ℎ=1∶𝑟do

            Parse MeSH heading ℎ into tokens

            Add each token 𝑡 with index 𝑝 to 𝑤^′_𝑗, the 𝑗th column of 𝑊^′

       i.e., 𝑊^′_𝑝𝑗=𝑊^′_𝑝𝑗+𝑔^′_𝑝×𝐻_𝑗𝑖

      end for

end for