Machine-learning classification with additivity and diverse multifractal pathways in multiplicativity

Evidence of multifractal structures has spread to a wider set of physiological time series supporting the intricate interplay of biological and psychological functioning. These dynamics manifest as random multiplicative cascades, embodying nonlinear relationships characterized by recurring division, branching, and aggregation processes implicating noise across successive generations. This investigation focuses on how well the diversity of multifractal properties can be specific to the type of cascade relationship between generation (i.e., multiplicative, additive, or a mixture) as well as to the type of noise (i.e., including additive white Gaussian noise, fractional Gaussian noise, and various amalgamations) among 15 distinct types of binomial cascade processes. Cross-correlation analysis of multifractal spectral features confirms that these features capture nuanced aspects of cascading processes with minimal redundancy. Principal component analysis using 13 distinct multifractal spectral features shows that different cascade processes can manifest multifractal evidence of nonlinearity for distinct reasons. This transparency of multifractal spectral features to underlying cascade dynamics becomes less amenable to machine-learning strategies. Fully connected neural networks struggled to classify the 15 distinct types of cascade processes based on the respective multifractal spectral features (<a:math xmlns:a="http://www.w3.org/1998/Math/MathML"><a:mrow><a:mn>45.5</a:mn><a:mo>%</a:mo></a:mrow></a:math> accuracy) yet demonstrated improved accuracy when addressing single categories of cross-generation relationships, that is, additive (<b:math xmlns:b="http://www.w3.org/1998/Math/MathML"><b:mrow><b:mn>91.6</b:mn><b:mo>%</b:mo></b:mrow></b:math>), multiplicative (<c:math xmlns:c="http://www.w3.org/1998/Math/MathML"><c:mrow><c:mn>75.4</c:mn><c:mo>%</c:mo></c:mrow></c:math>), or additomultiplicative (<d:math xmlns:d="http://www.w3.org/1998/Math/MathML"><d:mrow><d:mn>70.6</d:mn><d:mo>%</d:mo></d:mrow></d:math>). While traditional principal component analysis reveals distinct loadings attributed to individual noise processes, multiplicative relationships between generations effectively make the constituent noise processes less discernible to neural networks. Neural networks may lack sufficient hierarchical depth required to effectively distinguish among nonadditive cascading processes, recommending either elaborating multifractal geometry or using alternate architectures for machine-learning classification of cascades with multiplicative relationships. Published by the American Physical Society 2024