Date of Award:

8-2026

Document Type:

Dissertation

Degree Name:

Doctor of Philosophy (PhD)

Department:

Computer Science

Committee Chair(s)

Soukaina Filali Boubrahimi

Committee

Soukaina Filali Boubrahimi

Committee

Shah Muhammad Hamdi

Committee

Hamid Karimi

Committee

Kevin Moon

Committee

Shuhan Yuan

Abstract

Machine learning models are often used to classify signals collected over time, such as heart rhythms, movement recordings, industrial sensor measurements, and scientific observations. These models can be accurate, but they are often difficult to understand. Users may need to know not only what a model predicted, but also what would have needed to change for the model to reach a different decision.

This dissertation studies counterfactual explanations for time series data. A counterfactual explanation answers a “what-if” question. For example, if a model classifies a signal as one activity instead of another, the explanation shows how the signal would need to change for the model to predict the other activity.

Time series data are harder to explain than many other data types because the order and shape of the signal matter. Changing a single point may make a model change its prediction, but the resulting signal may no longer look realistic. A good explanation should change the model’s decision while preserving the structure of the original signal.

This dissertation develops these explanations in several steps. First, it uses shapelets, which are representative subsequences of class-specific patterns, to guide counterfactual changes toward meaningful regions of a signal. Because mining shapelets can be time-consuming and a fixed pattern may not match every input, the next method learns a saliency map that identifies the important time steps for each individual signal. This idea is then extended to multivariate time series, where the explanation must decide both which time steps matter and which measured variables carry the relevant evidence.

A bake-off study then compares representative existing time series counterfactual methods. This comparison shows that different methods make different trade-offs: some are better at changing the model’s prediction, some stay closer to the original signal, and some make fewer changes. The bake-off also shows that many major methods are largely guided by the current input, a nearest unlike example, or an optimization process for that specific case. When the guide example is poorly aligned with the original signal, or when the optimization changes the signal in an unnatural way, the resulting counterfactual may contain abrupt changes or lose temporal coherence.

This observation motivates the next method in the dissertation, which learns a reusable generation policy from many training instances rather than relying only on one guide example or one optimization process for each case. The bake-off case studies also highlight an important remaining challenge: a counterfactual can successfully change the model’s prediction and still look unnatural as a time series. The final method addresses this problem by using diffusion models to generate counterfactuals that better follow the structure of real time series.

Overall, the dissertation aims to make time series classifiers more transparent and trustworthy. It shows that useful explanations should not only change a model’s prediction, but should also remain close to the original signal and preserve realistic temporal behavior.

Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS

Copyright for this work is retained by the student. If you have any questions regarding the inclusion of this work in the Digital Commons, please email us at .