Robust Time Series Chains

[seanlaw]

I just came across this new paper by Jessica Lin on a more robust time series chain.

Not sure if it is relevant or how it compares with #211 especially if we are able to leverage our top-k code

[seanlaw]

The paper also states the following:

The most closely related works are the ones by Zhu et al. [5] and Imamura et al. [6]. Zhu et al. [5] first introduce the concept of time series chains. The work enforces every adjacent pair of subsequences in the chain to be the left and right nearest neighbors of each other, and reports the longest chain as the top chain in the time series. Although the concept is simple and intuitive, the bi-directional nearest neighbor requirement is shown to be too strict in many real-world applications, as the chain can easily break by data fluctuations and noise [6]. Imamura et al. [6] relax the bi-direction condition by only enforcing the single-directional left nearest neighbor requirement, and added a pre-defined angle constraint to guarantee the directionality of the discovered chains. The concept is shown to be more robust than [5] in some conditions, but a new angle parameter is introduced, and we have observed that in some real-world scenarios, no meaningful chains can be found no matter how we set this parameter.

This is mostly consistent with what @NimaSarajpoor observed in his earlier explorations