Growing Self-Organizing Maps for Multi-label Classification

• Pedro Henrique Casarotto UFSCar
• Ricardo Cerri USP

Resumo

In Machine Learning, multi-label classification is the problem of simultaneously classifying an instance into two or more labels. It is a challenging problem since each label has its specialty, and correlations between them must be considered. A Self-Organizing Map (SOM) is a Neural Network where neurons organized in a grid are tuned to represent the input instances in self-organization. After tuning, similar instances in the input space are mapped to closer neurons in the grid. SOMs have already been used for multi-label problems, obtaining competitive results with other methods. However, the static nature of their grid of neurons is a disadvantage since it is difficult to define the optimized grid size for each problem. The Growing Self-Organizing Maps (GSOM) extends the SOMs, allowing the network to grow during execution based on the data characteristics. This paper proposes a GSOM to predict multi-label data. The experiments showed that GSOM obtained better or more competitive results in most of the datasets investigated compared to SOM and had a competitive performance compared to other methods.