Estimación de atribuciones de características para la clasificación de series temporales: predicción supervisada de Gradientes Integrados según el paradigma de Aprendizaje Multitarea = Estimating feature attributions for time series classification: supervised prediction of Integrated Gradients under the Multitask Learning Paradigm
Year: 2024
Type: dissertation
Abstract: Neural Networks are increasingly being applied in high-stakes classification scenarios such as diagnosing a malfunctioning heart. Due to the inherent complexity and lack of interpretability of neural networks, a range of explainability methods have emerged, with feature attribution being one of the most prominent. Feature attribution assigns a relevance score to each timestep, providing insights into which parts of the time series contribute to the classification outcome. Hereby, the explanation is made after the classification. However, the Multitask Learning paradigm allows for the simultaneous training of a neural network to produce dual outputs, such as classification and its corresponding explanation. In this thesis, we explore the potential of predicting integrated gradient-based feature attributions within a Multitask Learning framework. Predicting these feature attributions alone offers significant efficiency gains, as it requires only a single feed-forward pass. Furthermore, the utilization of multitask learning and the sharing of a common feature space does not only amplifies these efficiency gains but also brings potential advantages such as improved generalization capabilities of the neural network. In our experimental exploratory analysis, we discovered that predicting integrated gradient-based feature attributions is feasible for time series data, evidenced by the moderate to high similarity between predicted and computed attributions. Additionally, we demonstrate that incorporating multitask learning does not compromise the accuracy of classification and prediction tasks. Furthermore, we introduce an iterative updating strategy, wherein we align the target features for the explanation module with the classifier’s calculated feature attributions during training. The outcomes indicate that the explanations predicted simultaneously exhibit moderate to high similarity with the classifier’s actual integrated gradient feature attributions.
Author: Bönte, Béla
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6990968991