One-year project within the Master's program of the National Research University Higher School of Economics "Machine Learning and Highly Loaded Systems".
The project has two objectives:
- to investigate possible multi-point-ahead prediction strategies in the presence of one or many time series in a dataset;
- to finalize an open-source time series analysis tool.
Since our project is aimed at developing a tool for time series forecasting, we needed to find data on which we could conveniently and clearly demonstrate certain features of the tool.
In the first phase of the project, we looked at quite a lot of public data, which can be categorized into two sources:
- Monash Time Series Repository (Godahewa (et al.), 2021):
- Datasets that are most commonly found in articles about state-of-the-art architectures.
- NOTE: The data in the Monash Repository is initially provided in .tsv format. There is a function in
data/data_formatting.pyto convert them to .csv format (source). There is also a function that translates datasets from the second item into the required format.
- NOTE: The data in the Monash Repository is initially provided in .tsv format. There is a function in
All datasets we consider can be found here.
The list and main characteristics of the datasets we reviewed can be found here. And the analysis of these datasets is in EDA/EDA_other_datasets.
Among these data we have selected ETT dataset, which "contains the 7 indicators of an electricity transformer in two years, including oil temperature, useful load, etc." and is convenient for analysis, as it contains rather diverse time series, containing both trend and seasonality, and also has rather frequent granularity (15 minutes). Its analysis is located separately in EDA/EDA_ETT.ipynb. If necessary, the use of other datasets is assumed (e.g., to consider series with pronounced patterns of seasonality).
The project can be divided into 4 parts according to the number of participants. Each of us has more or less independent from each other field of activity and tasks.
- ML-models and transformations (@laplan)
- Implement Baseline models in the model module.
- Implement ML models (linear regression, tree-based) in the model module.
- Model combination and model ensemble.
- Combining different transforms, storing parameters of transformers, implementing forward and backward (transform, inverse_transform).
- Optional: accounting for heuristics (not ML techniques).
- Neural Networks (@Einstein_30)
- Consider available architectures for solving time series forecasting problems
- Implement state-of-the-art neural network architectures in the model module and general learning pipelines.
- Hierarchical series and clustering (@s21_fernando)
- Review the literature on time series hierarchy (building a combination of different models at different levels of the hierarchy that differ in granularity).
- Implement as a wrapper (preset) on an existing library.
- Optional: partitioning of time series into useful and not useful.
- MLOps, library support, strategies (@elineii):
- Consolidating results and helping with implementation, collecting technical comments and fixing bugs, improving usability.
- Service schema and API (Streamlit).
- MLOps: poetry, pre-commit, actions, etc.
- Selection of optimal prediction horizon for recursive and direct strategy mixed with multi-input-multi-output.
- Kostromina Alina (@elineii)
- Larchenkov Mikhail (@laplan)
- Zuikova Olga (@s21_fernando)
- Karagodin Nikita (@Einstein_30)