Title: Methods for Transfer-Learning of Time Series Data

Duration: 01.06.2021-31.12.2025

Research Area: Mathematical Foundations and Statistical Learning

Analysis of time series data is a challenging topic of several subject areas, in particular if several parallel time series are used to predict trends. Typical real-world issues are (1) difficult discrimination between auto-, cross- and cross-lagged correlations, (2) unequal spacing of time points, (3) missing values and (4) low number of available data points. In this project, we analyze, compare and improve many prediction models based on time-series data. We will consider for example, non-linear autoregressive exogenous models (NARX), long-short term memory models (LSTM), gated recurrent unit networks (GRU) and mechanistically inspired neural network, the latter aiming at combining natural intelligence derived mechanistic models and artificial intelligence.

In particular, we analyze the embedding and learning characteristics of such networks. We also explore the relative importance of different transfer-learning when parametrizing the model for different situations.

Aims

In the project “Methods for Transfer-Learning of Time Series Data”, we have three major aims:

• To compare embedding characteristics of different prediction models
• To develop suitable learning frameworks of models including different transfer learning approaches
• To assess the relative importance of natural intelligence derived mechanistic assumptions for prediction purposes

Problem

Different methods and approaches for time-series data-based prediction models are proposed. These models are typically used in a heuristic framework, which is difficult to generalize to other situations. In particular, there is no general theory or even guideline which method / model to be used in which situation. One reason is that the theoretical aspects behind these models are under-investigated. We aim to improve the situation by analyzing embedding characteristics of these models to identify suitable application scenarios. Moreover, we will analyze the learning characteristics of these models to decide, which data sets are suitable for certain model applications. Finally, the relative importance of natural and artificial intelligences for model building is evaluated to assess, for example, how well knowledge-driven models perform in relation to phenomenologic predition modelling.

Practical example

We applied our methods to the modelling of individual patient data under chemotherapy. In particular, we modelled time series data of blood cells to predict toxicity in subsequent therapy cycles to support clinical decision-making.

Technology

• Different time-series models such as NARX and GRU
• Biologically inspired modelling
• Transfer-learning

Outlook

We expect high potential in combing natural and artificial intelligence for time series modelling. Thus, we will further work on combinations of knowledge-driven and phenomenological derived modelling frameworks.

Publications

• Steinacker M, Kheifetz Y, Scholz M. Individual modelling of haematotoxicity with NARX neural networks: A knowledge transfer approach. Heliyon. 2023 Jul 5;9(7):e17890. doi: 10.1016/j.heliyon.2023.e17890. PMID: 37483774; PMCID: PMC10362198.

Team

Lead

• Prof. Dr. Markus Scholz

Team Members

• Marie Steinacker
• Yuri Kheifetz