LSTM tools of the Matlab allow us to perform classification of the time-series. It is assumed, that different time-series can have different length, but they should have same interval of measurements (e.g., min., days, weeks, etc.). In real situations different time-series can have different times of measurement – e.g., one time-series is measured on times 1, 3, 7, 9 and other on the times 2, 6, 8. Certainly, we can perform interpolation and to transform our time-series to the time-series with same measurement times (e.g. 2, 5, 8), but may be you can propose more effective solution? Thanks a lot for your answers beforehand. Regards, Sergey.
Hi Sergey,
I recommend the following 2 papers that perfectly match your needs. Both are based on Neural ODEs by Chen et al., 2018:
- Lechner et al., " Learning Long-Term Dependencies in Irregularly-Sampled Time Series ", 2020 - https://arxiv.org/pdf/2006.04418v1.pdf
- Morrill et al., " Neural CDEs for Long Time Series via the Log-ODE Method ", 2020 - https://arxiv.org/pdf/2009.08295.pdf
Regards
Dear Stéphane Breton , thanks for answer. I will carefully read these articles. Regards, Sergey.
RNNs (LSTM and GRU in particular) are not aware of the "distance" between the "time-steps" or whatever it is in a modeled sequence. Thus, the only way to reinvent the wheel is to scatter the examples in a sequence proportionally to their "distances" in your time-series, leaving the other examples to be zeroed or of some none-coding token.
There are, however, ways for modeling time-series with varying "distances" between tokens. One way is to model it in a PDE-based or ODE-based models. Thanks to Stéphane Breton there are a couple of links above!
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series