S1 is a synthetic dataset created through https://github.com/misantamaria/deepvats-foundation/blob/master/nbs/_synthetic_data.ipynb to analyze the power of the embeding space to show time series segments. As the image above show, it is defined to have four clearly defined segments:

Segment	Start date	End date	Days	Hours	Minutes	Timestamps
$df_1$	2022-03-01 00:00	2022-03-08 11:59	7.5	180	10800	540
$df_2$	2022-03-08 12:00	2022-03-14 23:59	6.5	156	9360	465
$df_3$	2022-03-15 00:00	2022-03-24 23:59	10	240	1440	720
$df_4$	2022-03-25 00:00	2022-03-28 23:59	4	96	57600	288
Complete	2022-03-01 00:00	2022-03-28 23:59	28	672	40320	2013

In the original analysis, the masked autoencoder architecture showed a great performance on detecting all the segments showin gthe changepoints as extreme points in the clusters joining three big clusters between them:

This experimentation was done using, in inference, a window size w = 54, stride s = 2 and UMAP as algorithm to project the embeddings. As our previous scalability and stability analysis showed, it is a better option using UMAP followed by PCA in their GPU version. Thus, to test MOMENT on the detection of the segments of S1 we use w = 54, s = 2 and PCA followed by UMAP.

This are the images for S1 for MOMENT used for the research paper without the preprocessing.

The rest of images are saved at paper_imgs folder.