Right on Time: Revising Time Series Models by Constraining Their Explanations

Time series data is ubiquitous and often analyzed using neural models, which can be affected by spurious factors or shortcuts. This article introduces Right on Time (RioT), a method that uses feedback on model explanations to mitigate these shortcuts, improving model robustness and generalization. The article explores the challenges of shortcuts in time series data and demonstrates the effectiveness of RioT on various datasets, including a newly introduced real-world dataset with naturally occurring shortcuts. The method leverages traditional explanation methods like Integrated Gradients and incorporates feedback in both the time and frequency domains. The article also discusses the computational costs and sources of feedback, providing a comprehensive overview of the method's practical applications and limitations. By revising the model based on feedback, RioT significantly diminishes the influence of shortcuts, steering the model to align with the correct reasons. The article concludes by highlighting the potential of RioT for mitigating shortcuts in time series data and suggests future research directions.