---

name: specialized-modalities description: "Operate PyOD's time-series, graph, embedding/text/image/audio, multimodal, and torch-backed modality workflows with precise optional-backend guidance." disable-model-invocation: true metadata: disco-role: operating license: BSD 2-Clause

PyOD Specialized Modalities

Use this sub-skill when the user is working outside ordinary numeric tabular anomaly detection: time series, graphs, text/image/audio embeddings, multimodal fusion, or torch-backed neural detectors. Keep basic tabular fit/predict detector usage routed to classic-detectors, ADEngine/CLI/MCP lifecycle work to automated-lifecycle, and persistence/thresholding/combination operations to model-operations.

Quick route

• Time series arrays, windowing, transductive detectors, or TimeSeriesOD, KShape, MatrixProfile, SpectralResidual, SAND, LSTMAD, AnomalyTransformer: read time-series.md. Run scripts/time_series_smoke.py for a deterministic CPU-only sanity check.
• PyTorch Geometric node-anomaly workflows, Data(x, edge_index), graph detector selection, SCAN's structure-only path, or pyod[graph] installs: read graph.md.
• EmbeddingOD, MultiModalOD, text/image/audio presets, local/custom encoders, audio features, AudioAE, and modality fusion: read embedding-audio.md.
• Optional extras, torch/neural detector backend choices, credentials, CPU/GPU cautions, and install variants: read optional-backends.md. Run scripts/modality_backend_probe.py before giving install/remediation advice.
• Errors, shape mismatches, missing extras, transductive API surprises, graph-data failures, neural small-data issues, and credential/cache problems: read troubleshooting.md.

Operating rules

1. Verify the installed optional backend before recommending a modality workflow. Core PyOD is enough for CPU time-series detectors except LSTMAD and AnomalyTransformer; graph, embedding, OpenAI, HuggingFace image/text, audio, and most neural detectors need extras.
2. Do not claim GPU support or speed unless the current runtime has actually been probed. PyOD torch models can usually run on CPU; set explicit CPU devices for reproducible smoke tests when a class exposes a device parameter.
3. Treat all PyOD graph detectors and MatrixProfile as transductive: use decision_scores_ and labels_ after fit() rather than out-of-sample predict().
4. For embedding workflows, remember that EmbeddingOD defaults to detector LUNAR, which is torch-backed. If torch is absent, choose a non-torch detector such as KNN, LOF, ECOD, or IForest, or install the torch extra.
5. For neural/deep detectors on small data, reduce batch_size, epochs, model width/depth, and prefer deterministic CPU smoke checks before expensive runs.