A new Learning Model with Graph-based Time Variant Changes during a Pandemic Period

Covid-19 has become a worldwide pandemic showing multiple variants over time and regions. This phenomenon has been the subject of deep learning investigation. Although promising performance has been achieved, these models fail to track the temporal evolution of this virus due to the lack of time-varying features. We propose a model called TvDNN (Time-varying Deep-Neural Network) which includes time-variant parameters into the covid-19 sentiment analysis. Central entities (variants) were detected within the input space. These entities were featured by a high betweenness centrality and a provable time-variance of the Covid pandemic. We formalized the correspondence between the time-evolving state of the virus mutation, e.g., covid-19, and the corresponding tweets’ publication for an accurate sentiment analysis learning. It was found that TvDNN’s performance outperforms the reported models in literature on Covid-19 sentiment analysis by achieving (99.86%) accuracy, (98%) AUC/ROC. To conclude, characterizing the domain dynamically, e.g., changes with time, is the key to success. Consequently, the developed DNN with better performance. This should be generalized as that dynamic input changes are one of the possible types of domains in neural network development.