Topology of a neural network for multi-label classification

The objective of this work is to maximize the performance of a neural network that deals with multi-label classification, therefore it can associate several class labels to a sample. To achieve this, the network topology and therefore the structure has been modified: inserting / removing and exchanging the order of the various levels, modifying the value of the parameters linked to them (e.g. The number of hidden levels or the number of filters). This work proposes a new topology called LSTM_GRU composed by a composition of Long Short-Term Memory and Gated Recurrent Units trained with variants of the Adam optimization approach. The proposed neural network topology is also combined with Incorporating Multiple Clustering Centers (IMCC), which further boosts classification performance. Multiple experiments on twelve data sets representing a wide variety of multilabel tasks demonstrate the robustness of the ensemble proposed, which is shown to outperform the state-of-the-art.