Algorithms for context-aware forecasting

Date

2020-08

Journal Title

Journal ISSN

Volume Title

Publisher

Abstract

Forecasting has shown its importance for daily life and business decision making for many years. However, accurate forecasting is still a challenging problem that attracts researchers explore tirelessly. Beside the accuracy, forecasting algorithms are also required to cope with different settings related to the predictors, such as its availability or its relationship with the forecasted outcome. These requirements make context-aware forecasting become a promising solution. Context-Aware forecasting enhances not only the accuracy but also its ability to answer more complex problems of the forecasting algorithms.

This dissertation addresses forecasting related problems with different context settings to cope with the variety of information availabilities and different demands. Specifically, we present algorithms for context-aware forecasting in the following manners: 1) we introduce an algorithm to integrate various information for a cotton yield forecasting problem. In addition, the spatial feature is incorporated in the loss function to derive generalization of the model; 2) we present a solution for forecasting in the case of limited available information. We borrow the decomposition technique from signal processing to enrich the feature sets. Beyond that, the design of the forecasting model is selective to cope with the characteristics of feature sets which eventually enhances overall performance. 3) we discuss a forecasting algorithm in which the forecasting outcome is weighted by each input feature. Specifically, a sequence to sequence model is developed to forecast what people need in the event of natural disasters using weather and social media information. Furthermore, an attention mechanism is incorporated into the forecasting model to show the focus on the input portion that most explains the outcome. 4) we address the issues of using unstructured text in reasoning a predicted event. An unsupervised technique is proposed to normalize the unstructured text in order to provide a better understanding of the reason behind it.

The fundamental objective of this research has been to conduct a comprehensive study of problems and solutions in their context settings. The work also facilitates advances in text data reasoning, feature generation and integration techniques that ultimately help address issues in dealing with data-driven forecasting solutions.

Description

Keywords

Forecasting, Algorithms, Machine Learning, Data Science, Context-Aware

Citation