What do the authors want to do?
Given a time series and an opaque time series classifier, the authors want to find the minimum number of changes to be performed to the given time series in order to change the decision to another class.
They prove that the problem is -hard, showing how it’s reducible to the Hitting Set Problem, and provide two types of transformations to solve this.
The method proposed is model-dependent on the random shapelet forest model.
How do they achieve that in a nutshell?
They propose two types of time series tweaking:
- Reversible: where they want to find a series of transofmrations where each transformation can override any earlier transformation with .
- Irreversible: the same as reversible, but the transformations cannot be overwritten by other transformation, which causes the distance between the current time series and the transformed one to be monotonically increasing, allowing for early abandoning if the cost is above the best successful transformation.
They propose two greedy algorithms, one for reversible and one for irreversible tweakings.
Why is this paper useful?
Since the method is bounded to the Random shapelet forest classifier, it’s not so much useful. The interesting part is the fact that they reduced the problem to the Hitting Set Problem (if RSF is the classifier used), which may be useful as a start to generalize this method for all types of black-box classifiers.
What ideas they take from which other papers?
Has it been cited in recent works that enhance the ideas?
What concepts are unclear to you?
What are the performances?
Personal comment
Further Notes and Quotes
tags: paper counterfactual sequential-models reference: Karlsson et al. - Explainable time series tweaking via irreversible and reversible temporal transformations.pdf