Input seqence length discussion

[yueyangu]

The results of the previous models long-term forecasting performance are generally obtained with input seq len=96. As you said, input seq len is an important factor that could influence the performance. Is it biased to compare your fine-tuned seq len performance with the fixed seq len results of baselines?

[tianzhou2011]

We did not compare with the baseline using a fixed input sequence length of 96; instead, we compared with the best reported number in their original paper. For example, in the case of PatchTST, they used a significantly longer input length, and we were unable to surpass their performance. Generally, we did not intend to re-run the baseline method because it is highly likely that we do not possess the same tuning skills as the authors. Thus, any lower performance might not indicate a deficiency in the baseline's performance, but rather a reflection of our tuning capabilities. Additionally, some baselines utilize dynamic input lengths in their main tables to achieve optimal performance, which makes comparisons even more challenging. As a result, we chose to compare based on the best reported numbers in their original paper.

…

On Wed, Dec 11, 2024 at 7:06 AM Yueyan Gu ***@***.***> wrote: The results of the previous models long-term forecasting performance are generally obtained with input seq len=96. As you said, input seq len is an important factor that could influence the performance. Is it biased to compare your fine-tuned seq len performance with the fixed seq len results of baselines? — Reply to this email directly, view it on GitHub <#52>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB3JGO6VRJJNYABOJFSUVXL2E5XYJAVCNFSM6AAAAABTMHD6T2VHI2DSMVQWIX3LMV43ASLTON2WKOZSG4ZTCNBQGYYDQMQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>