Interactive Classification with Real-Time Contrastive Explanations
Abstract: We propose a framework in which users collaborate with machines to solve classification tasks, aided by contrastive explanations. Among these, counterfactual explanations stand out for their intuitiveness and effectiveness. However, long-standing challenges in counterfactual generation involve the efficiency of the search process, the likelihood of generated instances, their interpretability, and in some cases, the validity of the explanations themselves. In this work we address all these issues to present the first generative framework suited for real time explainable interactive classification. Our method leverages a label disentangled regularized autoencoder to achieve two complementary goals: generating high-quality instances and promoting label disentanglement to enable precise control over the decision boundary. By modeling the class-conditional data distribution, the framework avoids computationally expensive gradient-based optimizations, instead directly generating explanations based on the counterfactual distribution. A user study on a challenging human-machine classification task demonstrates the approach's effectiveness in enhancing human performance, emphasizing the importance of contrastive explanations.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Krikamol_Muandet1
Submission Number: 3847
Loading