Feature Extraction Based on Exponential-Weighted Higher-Order Local Auto-Correlation: An Approach to Improve Data Characterization

Abstract

Motivated by complex phenomena embedded into time series, this paper proposes EHLAC (Exponential-Weighted Higher-Order Local Auto-Correlation), an approach to extract features from dynamic data based on polynomial relations over time. The main idea for this new approach is to preprocess data in order to improve modeling performance of different techniques. EHLAC extends the traditional HLAC (Higher-Order Local Auto-Correlation), introducing non-linear transformations in terms of its integrals, what inhibits or highlights the influences of observations within the auto-correlation function, highlighting a wider gamut of data characteristics. This approach is evaluated in a song classification scenario, whose results evidence that EHLAC complements the set of attributes of HLAC and improves modeling performance.