Explaining Rocket time series classifier model

Numerous scientific and engineering disciplines, including health care, astronomy, and high-performance computing, deal with time series data. A recent trend is to use machine learning (ML) to process this complex data, and ML-based frameworks are beginning to play a crucial role in a wide range of applications. Most methods for time series classification that achieve state-of-the-art precision have high computational complexity, requiring substantial training time even for small datasets and being infeasible for large datasets. Convolutional kernels are a single, powerful resource that can capture many of the characteristics utilized by existing time series classification methods. Using random convolutional kernels in the Rocket model achieves cutting-edge accuracy at a fraction of the computational cost of existing methods. In recent years, there has been a greater emphasis on explaining the predictions made by black-box AI systems that process images and tabular data. Less emphasis has been placed on clarifying the predictions of ambiguous AI systems dealing with time-series data. In the case of Rocket, there are no explainability techniques available to help us comprehend the decision-making process of the model effectively without adding more complexity through the need to train another model to explain Rocket’s decisions. This thesis will develop a model-specific XAI technique in order to increase the transparency of Rocket’s predictions.