Hi everyone,
I am currently working on time series forecasting on quite sparse time series.
Problem: I am trying to predict sparse multivariate timeseries with hierarchical relations and graph like behaviours with only a small dataset (1month span).
Dataset: Some background on the type of dataset I work with :
The dataset is composed of events; each event is a flow of information between multiple applications, similar to a stack trace but at the application/enterprise level. These events have (reference_id, list[sources], list[targets], list[inforamtions], timestamp(s)).
- The dataset is quite small, about 1_000 h x ~500 categories (and growing)
- About 90% of these series are what I consider sparse (under 5events / hour)
What I tried: I have done preliminary work on predicting with univariate ML models (nixtla framework) but this led to only minor improvements from the naive approach.
What I have found so far:
- This discussion [Multivariate time-series transformer] about multivariate transformer-based models. Did not seem to lead.
- Graph Convolutional Neural Networks and such, but they often deal with years of data to back it up, and suppositions I cannot make.
- Time Series Transformer, yet I don’t see how I can overcome the (very) small dataset at hand.
- Transfer learning seems to be out of the question, as only multivariate models would accommodate the dependencies between series.
I hope I was clear enough with my problem at hand 
If you have any idea or know any related work that would be relevant, please share it! 