Rule-Extracted Neural Networks for Ethnomusicological Classification of Digitized Folk Music Archives

The pre-dominant approach for such situations is rule extraction post facto. However, simply extracting rules will not provide viable solutions since these rules only provide an approximate understanding of the underlying network. On our data, a TREPAN-style approach agrees with its own network on 95.4% of melodies, and there is no way to identify the remaining ones. Instead of extraction, we focus on learning rules directly. Differentiable Rule Networks (DRNs) provide differentiable rules. Therefore, learning rules directly results in the deployment of a rule-based system. Three variants of DRNs, where each component is altered, address the design costs. The Fuzzy Rule Network (FRN) grades rule satisfaction as opposed to binary satisfaction; the Neural Rule Ensemble (NRE) adds a sparse linear term; and the Motif-Rule Network (MRN) prepends a convolutional motif encoder over the input melody. The FRN yields a macro-F1 of 0.9615 ± 0.0054, using 13.6 rules, on 7,605 melodies from the Essen Folksong Collection, which are described by 34 musicological features. These printed rules exactly reproduce all test predictions, yielding a fidelity of 1.0000. The grading satisfaction of the rules yields an additional gain of +0.0195 ± 0.0060. We determined why no rule models achieve logistic regression (0.9802), which indicates that the tasks under consideration are nearly linearly separable. The negative results also explain why no rule models achieve logistic regression (0.9802): NRE retains no rules at all in four seeds of ten; and MRN's encoder adds +0.0051 ± 0.0088. In both cases, the features yield a linearly separable task. The rest of the best baseline models generalize well, with train – test gaps reaching +0.75.