README.md

metadata

license: apache-2.0
tags:
  - timeseries
configs:
  - config_name: default
    data_files:
      - split: gift_eval_256
        path: gift-eval_256/*.arrow
      - split: gift_eval_512
        path: gift-eval_512/*.arrow
      - split: electricity_256
        path: electricity_256/*.arrow
      - split: electricity_512
        path: electricity_512/*.arrow
      - split: ett1_15T_256
        path: ett1-15T_256/*.arrow
      - split: ett1_15T_512
        path: ett1-15T_512/*.arrow
      - split: rule_dataset_512
        path: rule_dataset_512/*.arrow
dataset_info:
  features:
    - name: target
      sequence: float64
    - name: dataset
      dtype: string
    - name: seq_length
      dtype: int64
  splits:
    - name: gift_eval_256
      num_examples: 11000
    - name: gift_eval_512
      num_examples: 11000
    - name: electricity_256
      num_examples: 10000
    - name: electricity_512
      num_examples: 10000
    - name: ett1_15T_256
      num_examples: 7000
    - name: ett1_15T_512
      num_examples: 7000
    - name: rule_dataset_512
      num_examples: 11000

Time-TSNE-Plot

A multi-scenario evaluation collection for t-SNE visualization of time-series embeddings. Each split stores samples with a fixed window length; the dataset field carries different label semantics depending on the split (see below).

Evaluation Dimensions

Dimension	Description	Recommended splits
1. Embeddings across datasets	Multiple real/benchmark datasets; labels are dataset names	gift_eval_256 / gift_eval_512
2. Embeddings across rule-based series	Rule generators such as triangle waves, square waves, ARMA, etc.	rule_dataset_512
3. Embeddings across channels (same dataset)	Multivariate ETT1 with sliding windows per channel; labels are channel indices	ett1_15T_256 / ett1_15T_512
4. Embeddings across time positions (same channel)	Univariate Electricity with sliding windows; labels are window start timestamp indices	electricity_256 / electricity_512

Splits Overview

GIFT-Eval (labeled by dataset name)

Sampled from GIFT-Eval subsets; dataset is the source dataset name (e.g. bitbrains_fast_storage, m4, etc.; 11 categories in total).

Split	Directory	# Samples	Window length
gift_eval_256	gift-eval_256/	11,000	256
gift_eval_512	gift-eval_512/	11,000	512

Electricity (labeled by timestamp)

Sliding-window segments on a univariate Electricity series; dataset is the start timestamp index of each window (string, e.g. "0", "1", …).

Split	Directory	# Samples	Window length
electricity_256	electricity_256/	10,000	256
electricity_512	electricity_512/	10,000	512

ETT1-15T (labeled by channel)

Sliding-window segments on multivariate ETT1 (15-minute) series; dataset is the channel index ("0"–"6", 7 channels in total).

Split	Directory	# Samples	Window length
ett1_15T_256	ett1-15T_256/	7,000	256
ett1_15T_512	ett1-15T_512/	7,000	512

Rule (labeled by generator)

Synthetic series from rule-based generators (e.g. generate_triangle_wave, generate_square_wave, generate_arma_samples, etc.; 10 categories); dataset is the rule/generator name. Only window length 512 is provided for now.

Split	Directory	# Samples	Window length
rule_dataset_512	rule_dataset_512/	11,000	512

Field Descriptions

Field	Type	Meaning
target	list[float64]	Time-series values within the window
dataset	string	Label (semantics vary by split; see above)
seq_length	int64	Window length (256 or 512)

Usage

from datasets import load_dataset

ds = load_dataset("whenxuan/Time-TSNE-Plot")

# 1) Cross-dataset embeddings
gift = ds["gift_eval_256"]

# 2) Rule-based series embeddings
rules = ds["rule_dataset_512"]

# 3) Different channels within the same dataset
ett = ds["ett1_15T_256"]

# 4) Different time positions within the same channel
elec = ds["electricity_256"]