Multivariate Time Series Data Causal Discovery

One of the goals for Artificial Intelligence is to achieve human-like intelligence. To that
\nend, several solutions were proposed over the decades, where causal structure discovery
\nwas proposed as a viable tool for enabling human-like reasoning. It can be treated as two
\nstages, first causal discovery that examines the cause-effect relationships between variables,
\nwhich are then used in the second stage, referred to as causal parameter inference, to
\nperform causal inference using counterfactual/logic-like reasoning similar to how human
\nbeings approach a problem. Generally speaking, there are two types of causal discovery
\nalgorithms: those that work with random variables and those that work with time series
\ndata. The focus of this thesis will be on the latter.
\nPerforming causal studies on real world dataset is very challenging for time series data
\nas it is prevalent to run into missing values. Currently, all existing causal algorithms require
\nevenly-sampled time series data which unfortunately are not always available.
\nIn this thesis I proposed a systems that can address this difficulties that is hindering
\ncausal learning on real world datasets. The proposed system performs causal discovery
\nusing time series data with missing entries (i.e., sparsely sampled data at varying intervals).
\nThe solution put forward for this task is comprised of two parts: data filling with Gaussian
\nProcess Regression, and causal learning using a the traditional Vector Autoregressive Model
\nor Machine Learning based approach. For the first part, experiments have shown that
\nGaussian Process Regression outperformed all the benchmark filling techniques such as
\nK Nearest Neighbour regression, Parametric Linear filling as well as random variable
\nfilling. The obtained Root Mean Square Error for GPR filled was the smallest under across
\nall filling percentages, comfortably beating benchmark algorithms by margins (RMSE
\ndifference varies from 0.05 to 1.5). As for the second part, an Echo State Network for
\ncausal learning is used due to its fast running time and higher prediction capabilities when
\ncompared with other causal learning algorithms available in the industry such as algorithms
\nlike Structural Expectation Maximization (SEM), and Subsampled Linear Auto-Regression
\nAbsolute Coefficients algorithm (SLARAC). When working with a 10 percent missing
\nentries, the proposed system is capable of obtaining an MCC score of 0.31 on a -1 to +1
\nscale where +1 represents perfect prediction and -1 represents complete no usefulness of
\nthe result. The MCC score received from the proposed system significantly outperformed
\nother methods such as SEM and SLARAC. To showcase the ability of the proposed system
\nto adapt causal relationships on real world engineering applications, the experiment was
\nconducted using a chemical refinery dataset called the Tennessee Eastman (TE) dataset.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.