Algorithm 3: LVW-MECO feature selection algorithm

Input: Dataset D containing N classes; feature set F; base classifier (binary classifier) algorithm h in OVO; aggregation strategy s in OVO; stopping condition control parameter T in M-LVW algorithm; and number of binary classifiers to be optimized r Initialization: number of binary classifiers $\mathrm{k}=\mathrm{N}(\mathrm{N}-1)/2$ Step 1: Apply the M-LVW algorithm to the k binary classifiers in the OVO setting individually and set the evaluation criterion for the parameter feature subsets as the accuracy of the corresponding binary classifier. This process selects k different feature subsets ${\mathrm{F}}_1,{\mathrm{F}}_2,\dots ,{\mathrm{F}}_{\mathrm{k}}$ for the k binary classifiers ${\mathrm{h}}_1,{\mathrm{h}}_2,\dots ,{\mathrm{h}}_{\mathrm{k}}$. Step 2: By applying the 10-fold cross-validation method, the F1 values corresponding to the k binary classifiers are calculated. All binary classifiers are then sorted in ascending order based on their F1 values. Step 3: For j = 1 to r /*Repeat the following Steps 4–5 sequentially for the r binary classifiers with lower F1 values.*/ Step 4: M-LVW algorithm is used for the ${\mathrm{p}}_{\mathrm{j}}$ binary classifier, and the evaluation criterion of its parameter feature subset is set as the F1 value of the binary classifier, and the feature subset ${\mathrm{F}}_{{\mathrm{p}}_{\mathrm{j}}}^{\prime }$ is selected for the ${\mathrm{p}}_{\mathrm{j}}$ binary classifier, and ${\mathrm{h}}_{{\mathrm{p}}_{\mathrm{j}}}^{\prime }$ is used as the feature set of the binary classifier e. Step 5: The k binary classifiers of ${\mathrm{h}}_1,{\mathrm{h}}_2,\dots ,{\mathrm{h}}_{\mathrm{k}}$ are used to form a multi-classifier H with aggregation strategy s; ${\mathrm{h}}_{{\mathrm{p}}_{\mathrm{j}}}^{\prime }$ and other binary classifiers ${\mathrm{h}}_{\mathrm{i}}$, $\mathrm{i}\ne {\mathrm{p}}_{\mathrm{j}}$, and a multi-classifier is formed by aggregation strategy s; then, the accuracy of H and ${\mathrm{H}}^{\prime }$ is obtained by using the 10-fold cross-verification method. If the accuracy of ${\mathrm{H}}^{\prime }$ is greater than H, then ${\mathrm{F}}_{{\mathrm{p}}_{\mathrm{j}}}^{}={\mathrm{F}}_{{\mathrm{p}}_{\mathrm{j}}}^{\prime }$, ${\mathrm{h}}_{{\mathrm{p}}_{\mathrm{j}}}^{}={\mathrm{h}}_{{\mathrm{p}}_{\mathrm{j}}}^{\prime }$. Output: The subset of features ${\mathrm{F}}_1,{\mathrm{F}}_2,\dots ,{\mathrm{F}}_k$ corresponding to each binary classifier in OVO