Decoding the scientific creative-ability of subjects using dual attention induced graph convolutional-capsule network

Abstract

There is an increasing demand of creative individuals in scientific research, innovation sectors of software industries and industrial research/development sectors. On the other hand, there are requirements of analytical minded people in investigation departments, academia and management sectors. Unfortunately, classification of individuals into creative and analytical categories based on their behavioral response is not easy. This paper makes an attempt to classify people into 4 categories: Analytical, High Creative, Medium Creative and Low Creative from their brain response during their participation in a creativity-test based on convergent problems. The proposed classification problem involves two main phases. In the first phase, a brain connectivity map is constructed from the electroencephalogram (EEG) response of the brain using Pearson’s correlation technique. In the second phase, a set of three centrality features, namely degree, closeness and betweenness are extracted from the connectivity map and fed to a classifier model for categorizing the afore-said class labels. The classifier model here synergistically combines one Graph Convolution Network (to abstract the brain connectivity-based centrality features) and one Capsule Network (to undertake the classification task) to develop the proposed Dual Attention Induced Graph Convolutional-Capsule network (DAIGC-CapsNet). The novelty of the proposed classifier model lies in the dual attention module and a new routing algorithm. The dual attention modules includes a) a Mish Induced Attention Module (MI-AM) to guide the graph convolution layers to focus on the most significant node attributes, and b) a Fused Attention Module (F-AM) to ensure the transmission of the most relevant predictions from the primary capsule to the class capsule layers. The latter attention module combines the effects of two sub-modules (channel and spatial) that concentrate on determining "what" and "where" to prioritize within the channel and spatial dimensions of the primary capsules. Lastly, the coupling between the primary and class capsule layers is strengthened by a Sparsemax based routing algorithm. Experiments conducted yield fruitful and definitive outcomes that substantiate the effectiveness of the proposed framework with respect to its conventional counterparts. Moreover, statistical validation of the proposed classifier using Friedman’s test also proves its efficacy compared to its competitors.