Dr Futoon Abu Shaqra  completed her PhD in April 2024 at RMIT University.

Thesis Title
Modeling Heterogeneous Time-series with Multi-resolution Sporadic Data.

Research Description
Futoon’s thesis revolutionizes the modeling of complex real-world time series data by developing streamlined processes to handle irregular, highly dimensional, and heterogeneous datasets. In a world where rapid technological advancements continually generate vast amounts of temporal data, particularly in IoT and smart city applications, the challenges of traditional time series analysis become evident. Systems like healthcare or environmental monitoring produce irregular, high-dimensional, and non-stationary data, making it difficult for conventional machine learning models to capture intricate relationships without extensive preprocessing. Unfortunately, such preprocessing not only consumes time and resources but also distorts temporal dependencies, affecting prediction accuracy and timely access to information.

Futoon’s thesis addresses these challenges with significant contributions in three primary areas. Firstly, she introduces the Parallelised Irregularity Encoders for Forecasting with Heterogeneous Time Series (PIETS) and PIETS+ frameworks. These novel approaches are tailored to handle the complexities of multi-source time series analysis, a common challenge in real-world applications. Unlike traditional methods, PIETS and PIETS+ effectively capture temporal patterns and correlations within features, enhancing predictive capabilities for both one and multi-step forecasting while modeling heterogeneous time series. By leveraging information from diverse data sources, these models not only excel in capturing temporal data complexity but also expedite the training process.

Moving on, Futoon delves into the challenge of irregularity, particularly focusing on highly sporadic time series with consecutive unobserved values. Traditional approaches struggle with such irregularities, leading Futoon to explore the performance of neural ordinary differential equation (ODE) models on sparsely observed time series data. Introducing SeqLink, a robust neural-ODE architecture for modeling partially observed time series, she demonstrates its ability to generate continuous representations regardless of sequence length or data sparsity. Unlike traditional methods relying solely on the last observed value, SeqLink utilizes ODE latent representations from multiple data points, enhancing sequence representation robustness.

Expanding her contribution further, Futoon tackles irregular streaming time series and continual learning, crucial for real-time applications. Existing continuous learning models often require buffering lengthy sequences, hindering responsiveness, and are typically tailored for regularly sampled sequences, which may not reflect real-world scenarios. In response, she introduces ODEStream: a Buffer-Free Online Learning Framework with an ODE-based Adaptor for Streaming Time Series Forecasting. This innovative approach adapts to data irregularity and concept drift without complex frameworks, mitigating performance degradation over time by learning sequence dynamics changes, providing a streamlined solution for real-time analysis

Supervisors
Prof Flora Salim, University of New South Wales
Dr Hao Xue, University of New South Wales
Dr Yongli Ren, RMIT University