An Ensemble Method for High-Dimensional Multilabel Data

Algorithm 1

Ensemble multilabel classifier using feature selection (EMCFS).

Input: 𝒟=𝐱₁,𝐲₁,…,𝐱_𝑛,𝐲_𝑛: The training multilabel data set;

    𝑟: The parameter of 𝑚_𝑘;

    𝑈=𝐱₁,…,𝐱_𝑠: The test data set without labels;

Output: 𝐘=𝐲₁,…,𝐲_𝑠: The label set of 𝑈;

Training stage

For each label 𝑙_𝑘 in 𝐿  do

  Obtain 𝑃𝑙_𝑘 and 𝑁𝑙_𝑘 of 𝑙_𝑘 from 𝒟 according to Def. (1) and (2);

  Calculate 𝑚_𝑘 with 𝑃𝑙_𝑘 and 𝑁𝑙_𝑘 according to Def. (6);

  For each feature 𝑎_𝑖 in 𝒟_𝑘=𝑃𝑙_𝑘∪𝑁𝑙_𝑘  do

    Calculate 𝜌^𝑘_𝑖+ and 𝜌^𝑘_𝑖− according to Def. (4) and (5);

  Select the features whose 𝜌^𝑘_𝑖+ or 𝜌^𝑘_𝑖− is larger than 𝑚_𝑘;

  Train a binary classifier 𝑓_𝑘 on 𝒟_𝑘 with the selected features;

Endfor

Prediction stage

For each sample 𝑜_𝑖 in 𝑈  do

  𝐲_𝑖=∅;

  For each classifier 𝑓_𝑘  do

     𝐲_𝑖=𝐲_𝑖∪𝑓_𝑘𝑜_𝑖, where 𝑓_𝑘 returns 0 or 1;

Endfor