|FALZONE EMANUELE||Cycle: XXXV |
Section: Computer Science and Engineering
Tutor: TANCA LETIZIA
Advisor: DELLA VALLE EMANUELE Major Research topic
:Composite Event ForecastingAbstract:
In many scientifics fields, measurements are collected over time, leading to the creation of time dependent dataset, i.e. time series. An extensive literature exists. It focuses on the identification of both descriptive models, whose aim is to describe the phenomena that is represented by the time series, and of predictive models, with the aim of forecasting future values of the time series. Many fields, such as economics and finance, base their business decisions on the accuracy of such models. As a consequence, in the last decades, a lot of effort has been put into enhancing predictive models, exploiting features related to the time series and correlation between multiple time series (multivariate time series). In many real scenarios time series can be naturally aggregated in multiple hierarchical structures, typically a directed acyclic graph, based on attributes such as geographical location, product type, etc. This allows for the possibility to look for both single and aggregated points along the hierarchy. For instance, manufacturing companies are interested in predicting the wholesales for each product. At the same time, in order to estimate raw materials and to plan deliveries, it is possible to exploit product aggregation based on material and country, respectively. Many approaches, such as top-down, bottom-up and optimal regression combination, have been proposed to enhance accuracy at different levels of the hierarchies. However, such approaches consider only one hierarchy (built on one feature) at a time. In addition, the complexity of hierarchies has a great impact on scalability. This thesis focuses on multivariate time series and aims at designing, implementing, and evaluating novel approaches that exploit multiple hierarchies to improve accuracy of existing predictive models. Our goal is to make such approaches scalable, so as to avoid performance problems when working with complex hierarchies.