Understanding the behaviour of machine learning models is helpful in several ways. For example, in gaining trust in model predictions, improving model architectures and identifying dataset faults. The existing approaches for analysing machine learning models, although helpful, have several limitations (e.g. they're application or model specific, computationally intensive, or non-robust). This work aims to address some of these challenges by developing explainability tools in the context of financial applications.
Explaining the science
Explainable machine learning involves (locally and globally) understanding the behaviour of machine learning models. An insight into model behaviour is essential for safety-critical applications (e.g. finance, healthcare). There exist two main approaches for analysing machine learning models. The first one involves developing inherently interpretable models (e.g. decision trees), while the second one involves using post-hoc methods to analyse pre-trained models.
Training inherently interpretable models is a promising research direction, but the majority of recent research in machine learning explainability focuses on developing post-hoc model analysis methods. Despite the progress in analysing machine learning models, several challenges remain. For example, the majority of explainability methods have been proposed and demonstrated in the context of image classification models and it is less evident how well the methods would generalise to models for other domains (e.g. finance). Moreover, recent works have questioned the reliability and robustness of some popular post-hoc explainability methods. Thus, it is essential to address these and several related challenges to develop explainability tools that can reliably explain the behaviour of machine learning models.
The project aims to develop reliable post-hoc model explainability methods that are applicable for different kinds of financial applications (that use structured or time series data). Additionally, the project will also involve developing inherently interpretable machine learning models (based on path signature features) that perform comparably to high performing black-box models (e.g. neural networks) in the context of financial time series applications.
The project aims to apply the proposed explainability tools to different kinds of financial applications. Some of the applications are: predicting the price movement of stocks using the limit order book data and predicting the probability of home equity line of credit default. Moreover, some of the developed methods would also be applicable to analyse machine learning models in other domains (e.g. computer vision, audio).