Temporal Information in Data Science: An Integrated Framework and its Applications

Data science is a well-known buzzword, that is in fact composed of two distinct keywords, i.e., data and science. Data itself is of great importance: each analysis task begins from a set of examples. Based on such a consideration, the present work starts with the analysis of a real case scenario, by considering the development of a data warehouse-based decision support system for an Italian contact center company. Then, relying on the information collected in the developed system, a set of machine learning-based analysis tasks have been developed to answer specific business questions, such as employee work anomaly detection and automatic call classification. Although such initial applications rely on already available algorithms, as we shall see, some clever analysis workflows had also to be developed. Afterwards, continuously driven by real data and real world applications, we turned ourselves to the question of how to handle temporal information within classical decision tree models. Our research brought us the development of J48SS, a decision tree induction algorithm based on Quinlan's C4.5 learner, which is capable of dealing with temporal (e.g., sequential and time series) as well as atemporal (such as numerical and categorical) data during the same execution cycle. The decision tree has been applied into some real world analysis tasks, proving its worthiness. A key characteristic of J48SS is its interpretability, an aspect that we specifically addressed through the study of an evolutionary-based decision tree pruning technique. Next, since a lot of work concerning the management of temporal information has already been done in automated reasoning and formal verification fields, a natural direction in which to proceed was that of investigating how such solutions may be combined with machine learning, following two main tracks. First, we show, through the development of an enriched decision tree capable of encoding temporal information by means of interval temporal logic formulas, how a machine learning algorithm can successfully exploit temporal logic to perform data analysis. Then, we focus on the opposite direction, i.e., that of employing machine learning techniques to generate temporal logic formulas, considering a natural language processing scenario. Finally, as a conclusive development, the architecture of a system is proposed, in which formal methods and machine learning techniques are seamlessly combined to perform anomaly detection and predictive maintenance tasks. Such an integration represents an original, thrilling research direction that may open up new ways of dealing with complex, real-world problems.