SDBM: Supervised Decision Boundary Maps for Machine Learning Classifiers

Understanding the decision boundaries of a machine learning classifier is key to gain insight on how classifiers work. Recently, a technique called Decision Boundary Map (DBM) was developed to enable the visualization of such boundaries by leveraging direct and inverse projections. However, DBM have scalability issues for creating fine-grained maps, and can generate results that are hard to interpret when the classification problem has many classes. In this paper we propose a new technique called Supervised Decision Boundary Maps (SDBM), which uses a supervised, GPU-accelerated projection technique that solves the original DBM shortcomings. We show through several experiments that SDBM generates results that are much easier to interpret when compared to DBM, is faster and easier to use, while still being generic enough to be used with any type of single-output classifier.

Automated summary

Understanding the decision boundaries of machine learning classifiers is crucial to gain insights into their functionality. Recently, a technique called Decision Boundary Map (DBM) was developed to visualize these boundaries using direct and inverse projections. However, DBM faces scalability issues when creating detailed maps and may generate results that are difficult to interpret when dealing with classification problems with multiple classes. In this paper, a new technique called Supervised Decision Boundary Maps (SDBM) is proposed, which addresses the limitations of DBM using a supervised, GPU-accelerated projection technique. Experimental results demonstrate that SDBM generates easily interpretable results compared to DBM, while also offering faster processing and user-friendliness. SDBM remains generic and can be utilized with any type of single-output classifier.