Interpreting time series forecasting models using model classreliance
Year: 2024
Type: conference-paper
Abstract: The widespread use of complex machine learning models in various data science tasks has made it difficult to understand how these models make decisions, leading to a growing interest in AI interpretability. This is especially true for time series data, where applying standard interpretability methods introduces unique challenges that have not been fully explored yet. This paper compares the model class reliance (MR) scores with existing state-of-the-art interpretability measures for time series forecasting models and provides the first application of MR for probabilistic time series forecasting. We employ several post-hoc interpretability techniques on different time series forecasting models and datasets, comparing MR scores against the benchmark SHAP scores. Additionally, we apply MR to deep probabilistic forecasting models, evaluating the outcomes against those achieved through feature removal and model retraining. Our findings indicate that MR scores are consistent with SHAP scores, particularly for important features, and can accurately prioritize feature importance in deep probabilistic forecasting models. Our work demonstrates that MR-based interpretations not only align with SHAP but also MR method stands out for its ease of implementation and efficiency, making it a viable option for interpreting complex time series forecasting models.
Authors: Sean Berry, Mucahit Cevik, Ozan Ozyegen
Institution: Toronto Metropolitan University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7126404571