A Full Explanation Facility for a MLP Network that Classifies Low-Back-Pain Patients and for Predicting its Reliability

This paper presents a full explanation facility that has been developed for any standard MLP network with binary input neurons that performs a classification task. The interpretation of any input case is represented by a non-linear ranked data relationship of key inputs, in both text and graphical forms. The knowledge that the MLP has learned is represented by average ranked class profiles or as a set of rules induced from all training cases. The full explanation facility discovers the MLP knowledge bounds as the hidden layer decision regions containing classified training examples. Novel inputs are detected when the input case is positioned in a decision region outside the knowledge bounds. Results using the facility are presented for a 48-dimensional real-world MLP that classifies low-back-pain patients. Using the full explanation facility, it is shown that the MLP preserves the continuity of the classifications in separate contiguous threads of decision regions across the 48-dimensional input space thereby demonstrating the consistency and predictability of the classifications within the knowledge bounds.