Layer 1: Text and Image Anomaly Detection

Full example: embedding_od_example.py

EmbeddingOD chains a foundation model encoder with any PyOD detector, enabling anomaly detection on text, image, and other non-tabular data. This implements the two-step approach (pretrained embeddings + classical OD) shown to outperform end-to-end methods in NLP-ADBench [BLLX+24].

1. Install the optional dependency and import.

   pip install sentence-transformers
   

   from pyod.models.embedding import EmbeddingOD
   

2. Prepare text data. Training data should represent the “normal” distribution. Anomalies are texts that differ in topic or style.

   train_texts = [
       "Quarterly revenue exceeded expectations by 12 percent",
       "The company announced a new product line for Q3",
       "Stock price remained stable after the earnings report",
       "Board of directors approved the annual dividend",
   ] * 40  # 160 training samples
   
   test_texts = [
       "Annual report shows strong financial performance",    # normal
       "Cost reduction strategy yielded positive results",     # normal
       "The volcano erupted covering the island in ash",       # anomaly
       "Alien signals detected by deep space telescope",       # anomaly
   ]
   

3. Initialize an pyod.models.embedding.EmbeddingOD detector, fit, and predict. The encoder converts text to embeddings; the detector finds outliers in the embedding space.

   clf = EmbeddingOD(encoder='all-MiniLM-L6-v2', detector='KNN',
                     contamination=0.1)
   clf.fit(train_texts)
   
   labels = clf.predict(test_texts)           # binary labels (0 or 1)
   scores = clf.decision_function(test_texts)  # anomaly scores
   proba = clf.predict_proba(test_texts)       # probability estimates
   

4. Alternatively, use a preset for one-line setup.

   # Fast text detection (MiniLM + KNN, no API key needed):
   clf = EmbeddingOD.for_text(quality='fast')
   

   Additional presets require extra packages:

   # Best text detection (requires: pip install openai, plus OPENAI_API_KEY):
   clf = EmbeddingOD.for_text(quality='best')
   
   # Image anomaly detection (requires: pip install transformers torch):
   clf = EmbeddingOD.for_image(quality='balanced')
   

5. Use a custom encoder function for any embedding source.

   import numpy as np
   
   def my_encoder(texts):
       # Your custom embedding logic here
       return np.random.randn(len(texts), 128)
   
   clf = EmbeddingOD(encoder=my_encoder, detector='LOF')
   clf.fit(train_texts)
   

References

[BKalayciE18]

İlker Kalaycı and Tuncay Ercan. Anomaly detection in wireless sensor networks data by using histogram based outlier score method. In 2018 2nd International Symposium on Multidisciplinary Studies and Innovative Technologies (ISMSIT), 1–6. IEEE, 2018.

[BLLX+24]

Yuangang Li, Jiaqi Li, Zhuo Xiao, Tiankai Yang, Yi Nian, Xiyang Hu, and Yue Zhao. Nlp-adbench: nlp anomaly detection benchmark. arXiv preprint arXiv:2412.04784, 2024.