I'm trying to teach myself how to apply clustering algorithms to time-series data. I've recently come across a paper (https://robjhyndman.com/papers/wang.pdf) about using time series characteristics as inputs for clustering. Specifically, this paper extracts characteristics such as seasonality, trend, autocorrealtion, skewness, etc. for a set of time series then builds a self-organizing map to separate those time series into clusters.

For right now I just want to play with the data generated in Wang and Hyndman's paper, which looks like the following. I have 14-time series, each with 400 observations each. From there, I obtained chracterstics from each of them using the method in the paper.

enter image description here

The table contains the most important characteristics for each of the 14 time series, rescaled so every data point is between 0 and 1. I was wondering what an appropriate way to cluster this data would be. The paper uses a self-organizing map (SOM), but I'm not sure if that's doable with R (I'm most comfortable in R, and would like to stick with it for the time being). Would hierarchical clustering be appropriate here?

My largest hangup at the moment is that I don't have a strong grasp on when to use a certain clustering method, or if clustering based on characteristics rather than actual data points makes one method (hierarchical, SOM, k-means, etc.) more appropriate than another. If there are any experts at clustering algorithms here, some guidance would be much appreciated!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.