A0745
Title: A time series approach to explainability for neural netswith applications to risk-management and fraud detection
Authors: Branka Hadji Misheva - BFH (Switzerland) [presenting]
Abstract: Artificial intelligence (AI) is creating one of the biggest revolutions across technology-driven application fields. The finance sector offers many opportunities for significant market innovation, yet the broad adoption of AI systems heavily relies on our trust in their outputs. Trust in technology is enabled by understanding the rationale behind the predictions made. In other words, it needs to ensure that values and domain knowledge are reflected in the algorithms' outcomes. To this end, the concept of eXplainable AI (XAI) emerged, introducing a suite of techniques attempting to explain to users how complex models arrived at a certain decision. Even though many of the classical XAI approaches can lead to valuable insights about the models' inner workings, in most cases, these techniques are not tailored for time series applications due to the presence of possibly complex and non-stationary dependence structure of the data. A generic XAI technique for deep learning methods (DL) is proposed, which preserves and exploits the natural time ordering of the data by introducing a family of so-called explainability (X-)functions. This concept bypasses severe identifiability issues, related among others, to profane numerical optimization problems, and it promotes transparency by means of intuitively appealing input-output relations ordered by time.
