Multi-label classification aims at assigning more than one label to each instance. Many real-world multi-label classification tasks are high dimensional, leading to reduced performance of traditional classifiers. Feature selection is a common approach to tackle this issue by choosing prominent Features. Multi-label Feature selection is an NP-hard approach, and so far, some swarm intelligence-based strategies and have been proposed to find a near optimal solution within a reasonable time. In this paper, a hybrid intelligence algorithm based on the binary algorithm of particle swarm optimization and a novel local search strategy has been proposed to select a set of prominent Features. To this aim, Features are divided into two categories based on the extension rate and the relationship between the output and the local search strategy to increase the convergence speed. The first group Features have more similarity to class and less similarity to other Features, and the second is redundant and less relevant Features. Accordingly, a local operator is added to the particle swarm optimization algorithm to reduce redundant Features and keep relevant ones among each solution. The aim of this operator leads to enhance the convergence speed of the proposed algorithm compared to other algorithms presented in this field. Evaluation of the proposed solution and the proposed statistical test shows that the proposed approach improves different classification criteria of multi-label classification and outperforms other methods in most cases. Also in cases where achieving higher accuracy is more important than time, it is more appropriate to use this method.

Feature selection is an important step in data preprocessing, which helps  reducing the dimensionality of data and simplifying the models. This process not only reduces the computational complexity of models, but also improves their accuracy by eliminating irrelevant Features and noise. The three most widely used approaches for Feature selection are filter, wrapper and embedded methods.  In this paper, first we review some support vector machine based Mixed-Integer Linear Programming (MILP) models and Supervised Infinite Feature selection (Inf-FS$_s$) method.  Then, we propose three hybrid approaches based on them. The first approach involves solving the relaxed linear model of the underlying  MILP model and then solving the MILP model for those Features with nonzero weights, namely a smaller MILP. In the second approach, first the Inf-FS$_s$ method is applied to rank the Features. Then depending on the Features costs, either chooses the top Features from the ranked Features until budget parameter is reached  or solves a knapsack problem to select cost effective Features. The third approach applies the first approach to the top $20\%$ of Features ranked by Inf-FS$_s$ method. To evaluate the proposed approaches' performance, experiments are conducted on four high-dimensional benchmark datasets for fixed and random Features costs. Results demonstrate that using either of the proposed approaches can significantly reduce running time of MILP models with comparable accuracies with the original MILP models.

Choosing a Feature vector for maximizing the success of a classifier machine is very effective. In this paper, using a combination of different methods to calculate the core function, an unsupervised Feature selection algorithm improvement has been proposed. Feature vector obtained by the proposed algorithm, will maximizes output accuracy of back propagation neural network classifier. In this paper we used case study of standard encoding of images compressed by alternate method and uncompressed images classifying based on their relative bit stream. Standards for classifications are JPEG and JPEG2000 and for uncompressed images is TIFF format. Using this Feature vector obtained by the proposed algorithm, classifier accuracy will be about 98%.