A Tsetlin Machine Framework for Universal Outlier and Novelty Detection

Outlier and novelty detection are two of the most active study areas where a huge amount of research effort has been made over the past decades. Although there are several well-known outlier and novelty detection methods, it is difficult to find one that can effectively and simultaneously deal with both tasks across different data types. When studied in detail, outliers and novelties exhibit different characteristics. In this paper, we introduce a universal Tsetlin Machine (TM) framework for novelty and outlier detection. The framework consists of a TM generator and a machine learning classifier. To this end, we enhance the vanilla TM with a generator to produce a novelty score. The generator consists of the conjunctive clauses of the TM, which are used to form a representative pattern of a given input. We demonstrate that the clauses provide a succinct interpretable description of the trained input and that our scoring mechanism enables us to discern outlier and novel input. Empirically, we evaluate our TM framework on nine outlier datasets, five novelty tasks, and a one-class classification setup. In all experiments, we were able to either outperform or closely match state-of-the-art methods, with the added benefit of an interpretable propositional logic-based representation.