Abstract
In virtue of omnipresence of temporal data (time series), a quest for efficient and interpretable models of such data becomes of paramount relevance. The accuracy-interpretable dilemma becomes more visible in light of the current trend of producing accurate, yet comprehensible models supporting users in making sound and highly actionable conclusions. This study is positioned in the realm of human-centric granular representation of temporal data (time series) where the logic settings of relational calculus is realized at the level of information granules. We establish modeling pursuits and resulting constructs at a certain level of abstraction delivered by information granules expressed as fuzzy sets. While have been some studies involving a framework of granular computing applied to granular models of time series, in this study we advocate a unified and original environment of logic blueprint of relational dependencies among information granules articulated by means of weighted logic expressions where both the components of the expressions as well as their associations are quantified (calibrated) in the presence of available experimental data. An overall processing scheme is structured as the following sequence of main functional modules: granulation (encoding) – relational computing – logic interpretation – numeric characterization of results (produced through decoding). Relational computing engages a suite of architectures that capture temporal dependencies describing predictions at the level of granular amplitudes and granular trends. The quality of the relational model is quantified at the level of information granules (which emphasizes the interpretability aspect) as well as. A series of experiments is reported to demonstrate the performance of the relational models and their performance impacted by the parameters of the relational architectures.