STR - Selective Temporal Replay

Code and results accompanying the paper "Bridging Continual Learning and Green Cloud Computing: Foundations for Sustainable Time Series Anomaly Detection".

STR is a continual-learning replay strategy for time-series anomaly detection: novelty-gated admission, retention-aware replacement, age-weighted preservation.

Table of contents

1. Repository layout
2. Environment setup
3. Data
4. Running experiments
5. Experiment outputs
6. License

---

1. Repository layout

STR/
├── config.py              Shared constants (seeds, batch sizes, data paths)
├── pyproject.toml         pdm project + dependencies
├── data/
│   ├── raw/{yahoo,smd}/   Raw benchmark data
│   └── consolidated/      Parquet versions of Yahoo (produced by the notebook)
├── notebooks/             Yahoo-data-consolidation.ipynb
├── scenarios/             Yahoo runner registry (DATASETS, MODELS, kwargs)
├── src/
│   ├── models/autoencoder/   GRU / LSTM / TCN × standard / variational
│   ├── regimes/              10 CL strategies (Naive, Cumulative, STE,
│   │                         EWC, LwF, RB, RES, GDumb, DER++, STR)
│   └── utils/                Data loading, reproducibility
├── scripts/
│   ├── run_experiments_yahoo.py   Yahoo!~A1/A3/A4 runner
│   └── run_experiments_smd.py     SMD runner
└── experiments/           Result CSVs + per-run artifacts (see §5)

2. Environment setup

Python 3.11 + pdm are required.

# install dependencies into a local .venv
pdm install

# (optional) activate the shell
$(pdm venv activate)

All commands below assume pdm run prefixes, which resolves to the project virtualenv automatically.

3. Data

Benchmark	Source	Used directly from
Yahoo!~A1, A2, A3, A4	data/raw/yahoo/ (CSV per series)	data/consolidated/yahoo_a{1..4}.parquet
SMD	data/raw/smd/ (machine-{1,2,3}-{1..N}.txt)	data/raw/smd/ (no consolidation step)

Yahoo - consolidation step (one-time)

Yahoo benchmarks ship as one CSV per series. The runner reads pre-consolidated parquet:

pdm run jupyter notebook notebooks/Yahoo-data-consolidation.ipynb

Execute all cells once to (re)generate data/consolidated/yahoo_a{1..4}.parquet. The parquet files are already in this repo, so this step is only needed if you re-extract the raw data or change the consolidation logic.

SMD - used as-is

SMD is multivariate (38 features) and read directly from data/raw/smd/. No consolidation needed.

4. Running experiments

There are two runners - one per benchmark family. Both are crash-safe (CSV is written after every run), resumable (--resume), and lightweight (~10 KB per run, no model weights persisted).

Runner	Benchmarks	Uses scenarios/?
scripts/run_experiments_yahoo.py	Yahoo!~A1, A3, A4	yes (registry-driven)
scripts/run_experiments_smd.py	SMD	no (self-contained configs)

Yahoo runner

# Yahoo A1 - full grid (10 strategies × 9 architectures × OFAT configs)
pdm run python scripts/run_experiments_yahoo.py --dataset yahoo-a1

# inspect the planned grid without running anything
pdm run python scripts/run_experiments_yahoo.py --dataset yahoo-a1 --dry-run

# resume after a crash / partial run
pdm run python scripts/run_experiments_yahoo.py --dataset yahoo-a1 --resume

# repeat for Yahoo A3 / A4
pdm run python scripts/run_experiments_yahoo.py --dataset yahoo-a3
pdm run python scripts/run_experiments_yahoo.py --dataset yahoo-a4

SMD runner

SMD partitions the 28 machines into 5 sequential concepts (cluster-based, domain-incremental):

C1 - cluster1 top-4   (1-6, 1-7, 1-1, 1-3)
C2 - cluster1 other-4 (1-2, 1-4, 1-5, 1-8)
C3 - cluster2 top-4   (2-2, 2-4, 2-9, 2-1)
C4 - cluster2 other-4 (2-5, 2-7, 2-6, 2-3)
C5 - cluster3 top-4   (3-8, 3-2, 3-10, 3-6)

# SMD - full grid
pdm run python scripts/run_experiments_smd.py

# dry-run / resume / single architecture / sanity-limit first N
pdm run python scripts/run_experiments_smd.py --dry-run
pdm run python scripts/run_experiments_smd.py --resume
pdm run python scripts/run_experiments_smd.py --model gru
pdm run python scripts/run_experiments_smd.py --limit 5

5. Experiment outputs

All runs land under experiments/{benchmark}/. The directory already contains the results used in the paper - re-running overwrites or appends via --resume.

Path	Size	Role
experiments/yahoo-a1/	74 MB	Primary Yahoo benchmark; main-text tables
experiments/yahoo-a3/	31 MB	Appendix tables
experiments/yahoo-a4/	31 MB	Appendix tables
experiments/smd/	96 MB	Primary SMD benchmark; main-text tables
experiments/yahoo-a1_buffer_profile/	3.8 MB	§5.7.3 buffer-profile tables + PCA figure
experiments/smd_buffer_profile/	181 MB	§5.7.3 buffer-profile tables + PCA figure

Per-benchmark structure

experiments/{benchmark}/
├── summary.csv                One row per run - all metrics, configs, CO₂
├── multiseed_summary.csv      (yahoo-a1, smd) extra seeds for headline cells
├── fwt.csv                    Forward-transfer values (recomputed)
└── {arch}/{run_id}/
    ├── params.json            Resolved hyperparameters for this run
    ├── codecarbon.csv         CodeCarbon emissions log
    └── *.parquet              Per-concept metric trajectories (for heatmaps)

summary.csv is the single source of truth for all paper tables and most figures. The per-run parquets feed the per-concept heatmap galleries.

6. License

MIT - see LICENSE.