HDC-ROCKETs

---

Table of Contents

• Overview
• Repository Structure
• Getting Started
  • Installation
  • Download the dataset
  • Running
  • Results
• License

---

Overview

Code to the paper [1]. The approach is based on the time series classification algorithm MiniROCKET [2], MultiROCKET [3] and HYDRA [4] and extend it with explicit time encoding by HDC.

[1] K. Schlegel, D. A. Rachkovskij, D. Kleyko, R. W. Gayler, P. Protzel, and P. Neubert, “Structured temporal representation in time series classification with ROCKETs and Hyperdimensional Computing,” Data Mining and Knowledge Discovery, 2025.

[2] A. Dempster, D. F. Schmidt, and G. I. Webb, “MiniRocket: A Very Fast (Almost) Deterministic Transform for Time Series Classification,” Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min., pp. 248–257, 2021.

[3] C. W. Tan, A. Dempster, C. Bergmeir, and G. I. Webb, “MultiRocket: multiple pooling operators and transformations for fast and effective time series classification,” Data Mining and Knowledge Discovery, 2022, doi: 10.1007/s10618-022-00844-1.

[4] A. Dempster, D. F. Schmidt, and G. I. Webb, “Hydra: competing convolutional kernels for fast and accurate time series classification,” Data Mining and Knowledge Discovery, vol. 37, no. 5, pp. 1779–1805, 2023, doi: 10.1007/s10618-023-00939-3.

---

Repository Structure

└── /
    ├── LICENSE
    ├── README.md
    ├── requirements.txt
    ├── configs/
    │   ├── defaults.yaml
    │   ├── HYDRA.yaml
    │   ├── MINIROCKET.yaml
    │   ├── MULTIROCKET.yaml
    │   ├── MULTIROCKET_HYDRA.yaml
    ├── create_figures_tables/
    │   ├── cd_diagram.py
    │   ├── create_eval_data.py
    │   ├── dataset_shapes_UCR_NEW.csv
    │   ├── figure_sec_4_3.py
    │   ├── figures_sec_5_2.py
    │   ├── figures_sec_5_3.py
    │   ├── figures_sec_5_3_example_data.py
    │   ├── figures_sec_5_56.py
    │   ├── plot_config.py
    │   ├── plot_ramifications.py
    │   ├── tables/         
    │   └── images/         
    ├── data/
    │   ├── constants.py
    │   └── dataset_utils.py
    ├── experimental_runs/
    │   ├── runs_hydra.sh
    │   ├── runs_minirocket.sh
    │   ├── runs_multirocket.sh
    │   ├── runs_multirocket_hydra.sh
    ├── models/
    │   ├── HYDRA_utils/
    │   ├── Minirocket_utils/
    │   ├── Multirocket_utils/
    │   ├── Multirocket_HYDRA/
    │   ├── hdc_utils.py
    │   ├── fit_beta_utils.py
    │   ├── Model_Pipeline.py
    │   └── model.pth
    ├── main.py
    ├── net_trail.py
    ├── results/

---

Parameter Description

The table below provides an overview of the main configuration parameters defined in configs/defaults.yaml. These parameters control dataset loading, encoding, model settings, and experimental execution.

Category	Parameter	Description
Dataset and Execution	dataset	Dataset name to use (UCR, UCR_NEW, synthetic, etc.)
	dataset_idx	Index of dataset to use (for ensemble-style runs)
	complete_UCR	Whether to run all datasets in the UCR collection
	complete_UCR_NEW	Whether to run all datasets in the UCR_NEW collection
	complete_UEA	Whether to run all datasets in the UEA collection
	hard_case	Use challenging version of synthetic dataset
	dataset_path	Path to dataset folder
	results_path	Path to output results
	log_level	Logging level (e.g., INFO)
General Parameters	model	Model type to run (MINIROCKET, MULTIROCKET, etc.)
	variant	Variant of the model (e.g., orig, hdc_oracle, hdc_auto)
	seed	Random seed for reproducibility
	seed_for_HDC	Whether to seed HDC vector generation
	seed_for_splits	Whether to seed data splits
	stat_iterations	Number of iterations (e.g., seeds) to evaluate
	n_time_measures	How many times to repeat timing for computational runtime
	batch_size	Evaluation batch size
	max_jobs	Maximum parallel jobs for multiprocessing
	vsa	Vector Symbolic Architecture (MAP)
	fpe_method	Fractional Power Encoding method (sinusoid)
	beta	Temporal encoding beta value
	multi_beta	Run multiple beta values at once
	best_beta	Automatically select best beta based on grid search
	kernels	Number of kernels to use
	nan_to_zero	Replace NaNs with zero in data
Normalization	normalize_input	Enable input normalization
	predictors_min_max_norm	Min-max normalize predictors
	predictors_z_score_with_mean	Apply mean normalization
	predictors_z_score_with_std	Apply std normalization
	predictors_norm	Apply vector length normalization
	norm_test_individually	Normalize test samples independently
MultiROCKET	multi_dim	Output dimension for MultiROCKET
	predictors_sparse_scaler	Use sparse scaler for predictors (if available)
Hydra	HDC_dim_hydra	Dimensionality used in HYDRA variants

---

---

Getting Started

Dependencies

Dependencies are listed in the requirements.txt file.

Installation

We recommend a virtual environment to run the code and using Python 3.10.

1. Clone the . repository:

git clone https://github.com/scken/HDC_ROCKETS.git

1. Change to the project directory:

cd HDC_ROCKETS

1. Install the dependencies: We recommend using mamba (conda) as virtual environment to run the code and using Python 3.10.

mamba create -n hdc-rockets python=3.10
mamba activate hdc-rockets
mamba install numpy=1.26 scipy=1.10.0 pandas=2.0.3 matplotlib seaborn=0.13.2 scikit-learn=1.5.2 sktime=0.34.0 aeon=0.11.1 h5py
pip install multi-comp-matrix rocket-fft tsai torch-hd pytorch-lightning hydra-core==1.3.2 tables

Download the dataset:

You can download the datasets and resample indices in two ways:

1. Official TSML Benchmark Repository
   Follow the instructions at https://tsml-eval.readthedocs.io/en/latest/publications/2023/tsc_bakeoff/tsc_bakeoff_2023.html to download the full benchmark datasets (112 + 30) and the corresponding resample indices.

2. Direct Download from Our Server
   Alternatively, you can download a prepackaged version directly from our cloud server:
   https://tuc.cloud/index.php/s/z5kZKSxose35sdK

wget https://tuc.cloud/index.php/s/z5kZKSxose35sdK/download -O dataset.zip
unzip dataset.zip -d data
rm dataset.zip

After downloading, extract and store all files in a single folder, e.g., data/.

Running:

• main.py is reading the config files specified in "configs" folder contains more specific parameters for running the experiment (default and model-specific ones)

Arguments to run

• all parameters which are not specified in the command line will be taken from the config files
• to overwrite a parameter from the config file, specify it in the command line (e.g. beta=0) or in the files itself

UCR Datasets:

UCR / NEW-UCR: single dataset vs full benchmark & original vs HDC variants

The extended UCR archive with the datasets from UCR plus 30 new, as explained in the paper, is indicated as UCR_NEW.

Variants:

• variant=orig → Original MiniROCKET (no HDC temporal encoding).
• variant=hdc_oracle → HDC‑MiniROCKET, evaluates a grid of betas and reports the best per dataset ("oracle").
• variant=hdc_auto → HDC‑MiniROCKET, automatically selects the best beta during the run.

Scope:

• Single dataset → set dataset_idx=<i>
• Full UCR benchmark → set complete_UCR_NEW=True

Run one dataset with the original MiniROCKET (here: dataset index 0):

python3 main.py --config-name=MINIROCKET variant=orig dataset_idx=0

Run one dataset with HDC‑MiniROCKET (auto beta):

python3 main.py --config-name=MINIROCKET variant=hdc_auto dataset_idx=0

Run the full UCR benchmark with HDC‑MiniROCKET (oracle betas):

python3 main.py --config-name=MINIROCKET variant=hdc_oracle complete_UCR=True

Run the full UCR benchmark with HDC‑MiniROCKET (auto beta):

python3 main.py --config-name=MINIROCKET variant=hdc_auto complete_UCR=True

Note: If you are using the NEW‑UCR archive (default in configs/defaults.yaml), add dataset=UCR_NEW to the command line to be explicit, e.g.:

python3 main.py --config-name=MINIROCKET variant=hdc_auto dataset=UCR_NEW complete_UCR_NEW=True

Synthetic Dataset:

• run normal synthetic dataset 1

python3 main.py --config-name=MINIROCKET variant=orig dataset=synthetic1

• run hard synthetic dataset 1

python3 main.py --config-name=MINIROCKET variant=orig dataset=synthetic1 hard_case=True

Multiple experimental runs:

To run multiple configurations of the model, use the scripts in the experimental_runs folder. The scripts are used to run multiple configurations of the model on different datasets. They include options for synthetic datasets and the UCR dataset, with different configurations for each.

e.g.:

sh ./experimental_runs/runs_minirocket.sh

• this will run all synthetic datasets and UCR in different model configurations of MiniROCKET

Results:

• the results will be written and saved in /results
• for each run a new folder will be created
• within this folder the results will be saved in from of Excel spreadsheets, text files for logging and json file for storing the used hyperparameters
• in addition, the code will copy all Python files to the result folder "Code" for reproducibility (snapshot of the code)

Creating Figures and Tables

The folder create_figures_tables/ contains scripts for visualizing results and generating LaTeX tables. For example, run the following to generate synthetic result plots:

python3 ./create_figures_tables/figures_sec_5_2.py

---

License

This project is licensed under the GNU General Public License v3.0.
See the LICENSE file for details.

↑ Return