Question about the EOS Token in Chronos

[Ilyaasshousha]

Does the EOS token appended to the quantized time series implicitly summarize the preceding sequence (since it "sees" the sequence during training), or is it purely structural with no content-related role? I ask because I aim to use this for clustering time series in an unsupervised manner.

Thank you!

[abdulfatir]

@Ilyaasshousha Yes, one can view the EOS token as a sort of global representative of the time series although that wasn't the original motivation of the token. It was only supposed to be an end of sequence marker. If you're looking for advice on which token's representation from the encoder to use for other time series tasks (clustering, classification, etc.), I would try the following two options:

• Using only the representation from the EOS token.
• Averaging representations across all tokens.

One of these may work better than the other depending on the task. I would also advise trying the embeddings from Chronos-Bolt models which should be much faster than Chronos.

I would be curios to know the results that you obtain with the two approaches.