Optimized Functional Link Artificial Neural Network for Multi-label Classification Anwesha Law 1 , Balasubramanyam Evani 2 , and Ashish Ghosh 1 1 Indian Statistical Institute, Kolkata, India {anweshalaw r, ash}@isical.ac.in 2 Manipal University Jaipur, Jaipur, India balasubramanyam.evani@gmail.com Abstract. Multi-label classification problem is a generalization of the traditional single-label classification, which makes the data more complex in nature. To handle the inherent complexity of multi-label data, a compact and efficient network known as functional link artificial neural network (FLANN) has been explored. FLANNs are known to functionally transform the input space to introduce non-linearity into the data, thus making the task of separating the classes in the output space comparatively simpler. In this paper, six multi-label FLANN models (five novel and one existing) have been devised for multi-label classification to procure the optimal configuration. These six variations of the network have been built using three basis functions - trigonometric, Chebychev, power polynomial and two learning techniques - backpropagation and particle swarm optimization. These fundamentals of FLANN have been thoroughly explored in the single-label domain, but are yet to be experimented for multi-label data. These multi-label FLANN models were tested on four datasets for ten performance metrics. Analysis of the results have generated some interesting conclusions and optimal models for multi-label classification. Keywords: Multi-label classification functional link artificial neural networks particle swarm optimization 1 Introduction Classification is a predictive data mining task, usually conducted by means of supervised learning. Traditional classification problems are annotated with a single-label, however, in the real world, data is mostly tagged with multiple labels. Multi-label classification, for example, is the task of tagging music with multiple emotions, catego- rizing movies into more than one genres and so on. Basically, the aim of multi-label classification is to predict a set of labels for each data instance as opposed to single-label classification where each input instance is associated with a singular class or label. Multi-label classification can be seen as a generalization of single-label classification. The methods to deal with the generality of multi-label data to produce the overlapping and convoluted boundaries can be broadly divided into - data transformation and problem adaptation based methods [9]. Most of the recent works [16, 17] in the past decade has been focused towards the problem adaptation based methods. This branch focuses on modifying the existing classification algorithms for multi-label data. While solving the problem of classifi- cation, artificial neural networks have shown the capability of solving very complex, non-linear data relationships [8]. The flexibility exhibited by neural networks make it capable of learning any type of pattern. They are loosely in- spired by how the human brain works; specifically, its interconnected neuron structure and its activation resembles the biological synaptic connections and its firing. This makes it capable of handling complex classification tasks. However, in literature, the application of neural network models for multi-label classification is somewhat limited. Adaptations of multi-layer perceptron (MLP) [17], radial basis networks (RBF) [16], functional link artificial neural networks (FLANN) [10], extreme learning machines (ELM) [11], etc. as multi-label classifiers have been shown in the literature. There is still scope of exploring this area with the problem of multi-label classification in mind. Among the various networks explored, FLANN is one such network that is compact yet efficient. Multiple FLANN adaptations for single-label classification [3,4,12] exist in literature, but it is yet to be sufficiently experimented in the multi-label domain. FLANNs are feed forward networks where the use of a hidden layer is omitted by non-linearly transforming the input features with some basis functions. The expanded input layer portrays a higher dimension projection of the input with better discriminating characteristics. Multi-label data is inherently quite complex in nature, mostly due to its multiple overlapping class boundaries. This calls for models that can handle this bottleneck and improve on it. Therefore, FLANN is an apt choice in this scenario. In the present paper, various models of FLANN have been devised in order to analyse each of their importance as a multi-label classifier. Three types of functional expansions viz. trigonometric, Chebychev and power polynomial expansions and two weight optimization techniques, namely, backpropagation and particle swarm optimization (PSO) have been included in this paper. 56 ICONIP2019 Proceedings Australian Journal of Intelligent Information Processing Systems Volume 16, No. 4