DeepVix: Explaining Long Short-Term Memory Network With High Dimensional Time Series Data

Machine learning automates the process of analytical model building by means of the computing power of machines. Visual analytics couples interactive visual representations and underlying analysis, putting the human at the center of the analytics and decisionmaking process. This paper aims to combine the strengths of both data science fields into a unified system, called DeepVix, which focuses on the visual explainability of the multivariate time-series predictions using neural networks. Within our DeepVix system, a visual presentation of the neural network explains the intermediate steps, as well as the temporal weights of various gates of the entire learning process. The relationships between input variables and the target variable can also be inferred automatically from the trained model. Interactive operations allow users to explore the neural network, to gain understandings of the model and essential features with layers and nodes, and finally to customize the neural network configurations to fit their needs. We demonstrate our approach with Recurrent Deep Learning on various real-world time series datasets, including the multivariate measurements of a medium-size High-Performance Computing Center, the S&P500 stock data over the past 39 years, and the US employment data retrieved from the Bureau of Labor and Statistics.