Analyse Interpr ´etable et Causale pour des S´ eries Temporelles Multivariées
Amin Dhaou · theses.fr (ABES) · 2024
Advances in artificial intelligence have led to the development of increasingly complex models for solving a wide range of tasks. In critical applications such as industry and medicine, it has become necessary to propose "interpretable" models that clearly establish the decision-making process, thus promoting understanding of these models and their decisions and, consequently, their user acceptance. These objectives fall within the field of eXplainable Artificial Intelligence (XAI), which has been attracting growing interest in recent years.Time-series data, which measure the evolution of variables over time, such as sensor readings or data monitoring, provide valuable information on the system's behavior. By identifying patterns in these data, we can understand the interactions between variables, improve forecasting accuracy, and design better intervention strategies. This thesis studies the analysis of high-dimensional time-series data, focusing on explaining local system deviations from normal operation and, on the global scale, modeling the underlying dynamics of the system to predict its evolution.This work has two main objectives. The first objective is to develop an interpretable algorithm that identifies the root causes of both normal and abnormal behavior in time series data. Various techniques are used to identify root causes, but they suffer from limitations in their ability to handle high dimensions and to distinguish causality from correlations.To overcome these limitations, an approach based on the concept of Granger causality [Granger 1988], which extracts interpretable and causal relationships in the form of rules, has been developed. The resulting algorithm is designed to handle different data types (numerical, categorical), provide users with interpretable explanations of the problem, and develop predictive rules to defuse the event in advance.The second objective aims to develop a forecasting model that not only predicts future values but also reveals the underlying dynamic of the time series influencing those predictions. This field, called symbolic regression, fosters transparency for users by explaining the model's reasoning. Regression models with parsimonious penalization are widely used in this field for their ability to learn complex dynamics in high-dimensional settings. Nevertheless, their forecasting performances can be limited, especially for complex, non-linear data. To address this, we propose a novel approach that combines penalized regression with forecasting error correction within a time series forecasting framework for improved learning of underlying dynamics.By achieving these goals, this research has the potential to significantly improve our ability to analyze and understand time series data. This will result in better forecasts, a better understanding of the system, and the development of more effective intervention strategies.