From Log Files to Prediction-Oriented Multidimensional Analysis: conceptual modelling of complex information to support advanced data analytics

Contemporary information systems generate large volumes of data whose structures, semantics, and temporal characteristics are heterogeneous and often only implicitly defined. Event logs capture detailed evidence of operational and business-process execution, but their diversity of formats, vocabularies, nested structures, and execution contexts hinders consistent interpretation, validation, and integration. In parallel, multidimensional analytical environments provide mature support for descriptive analysis, yet their conceptual models typically treat time as a single, uniform role and therefore do not make explicit the distinct temporal assumptions required when preparing data for prediction-oriented tasks. This thesis investigates how conceptual modelling can be extended to make structural, semantic, and temporal assumptions explicit in these two complementary settings. Following a design-science perspective, the work defines, formalises, implements, and evaluates conceptual artefacts through case-based and implementation-oriented evidence. The first contribution is a UML-based conceptual modelling approach for heterogeneous system event logs. It introduces log-specific modelling constructs, supports nested objects and semantic dependencies, aligns terminology with the Elastic Common Schema while enabling domain-specific extensions, and specifies mappings between UML models and JSON log structures. A supporting modelling tool demonstrates bidirectional transformations (forward generation and reverse engineering) together with schema-oriented validation. The second contribution is a prediction-oriented framework for multidimensional analysis that distinguishes three temporal roles: an Observation Window for feature construction, a Waiting Window for outcome latency, and a Prediction Window for target definition. The framework is formalised via Prediction-Oriented OLAP operators and a Prediction-Oriented Dimensional Fact Model that adds explicit temporal bindings and predictive dependencies. Python- and SQL-based implementations demonstrate how these temporal semantics can be operationalised in the construction of Analytical Base Tables. The evaluation provides evidence of implementability, reproducibility, and enforcement of the specified temporal constraints in the examined configurations, while not aiming to compare predictive algorithms or claim superior predictive accuracy.