Explainability aware clustering: a case study using COVID-19 analysis

Abstract

As the size of the time-series data increases, explaining large number of analyses becomes a critical issue. Interpreting time-series data with statistical summaries does not reflect the overall direction or the relationships between different time stamps in the time-series data. Visualizing multiple time-series data simultaneously also makes going through all possible time-series is still overwhelming for analysts. Machine learning models are effective in condensing information and have recently become more explainable. We utilize model explainability techniques to explain data analyses at a high level. Our approach is different than the traditional machine learning process where instead of making predictions on test data, we go back to the training data to understand it better by condensing the information in an explainable model. We propose a hybrid method that combines machine learning explainability with conventional data mining techniques. One problem is that, clustering similar time-series trends does not consider the explainability of the machine learning model. We apply state-of-the-art multi-objective clustering techniques to identify cohesive and explainable clusters. We demonstrate that pareto-based clustering techniques generate the most diverse cluster sets. We analyze the decision trees to identify the demographics that corresponds to each cluster.