{
  "field_id": "F018",
  "field": "anomaly detection",
  "directions": [
    {
      "direction_id": "G0",
      "name": "Normal-content restoration for anomaly detection",
      "scope": "Obtain a normal counterpart of an observation through reconstruction, completion, or denoising, and detect or localize anomalies from the differences."
    },
    {
      "direction_id": "G1",
      "name": "Temporal expectation and pattern deviations",
      "scope": "Use expected states of temporal behavior or observed recurring patterns as references to identify anomalous periods, trajectories, or subsequent observations."
    },
    {
      "direction_id": "G2",
      "name": "Distributional and sample-geometric outlier scoring",
      "scope": "Directly assign anomaly scores based on probabilistic evidence for observations under a normal distribution, or on their sparse or isolated positions relative to other samples."
    },
    {
      "direction_id": "G3",
      "name": "Anomaly-discriminative feature learning and references",
      "scope": "Make a feature space distinguish normal structure from deviations, and use normal feature references or learned representations for detection and localization."
    },
    {
      "direction_id": "G4",
      "name": "Learning with limited anomaly evidence",
      "scope": "When real anomaly examples are absent or scarce, construct surrogate anomalies, use a few known anomalies, or selectively obtain expert feedback to learn more targeted detectors."
    },
    {
      "direction_id": "G5",
      "name": "Structural and relational anomaly detection",
      "scope": "Use relationships among entities, attributes, or concurrently varying variables to determine whether structural or relational behavior deviates from normal, and localize the deviation where applicable."
    },
    {
      "direction_id": "G6",
      "name": "Video behavior anomaly recognition and localization",
      "scope": "Recognize anomalous behavior from actions, appearance, scenes, and their spatiotemporal relationships in video, and identify anomalous segments when needed."
    },
    {
      "direction_id": "G7",
      "name": "Anomaly capability transfer across target settings",
      "scope": "When target-specific anomaly training is limited or source and target conditions differ, transfer existing knowledge to maintain anomaly recognition and localization capability in the target setting."
    },
    {
      "direction_id": "G8",
      "name": "Sustained anomaly monitoring under deployment constraints",
      "scope": "Maintain timely and usable anomaly monitoring on continuously arriving data despite changing conditions or computational, storage, or device constraints."
    },
    {
      "direction_id": "G9",
      "name": "Anomaly explanation and attribution",
      "scope": "Provide usable feature attributions, explanations of causes, or counterfactual explanations of what would make flagged anomalies normal."
    },
    {
      "direction_id": "G10",
      "name": "Detector configuration and combination",
      "scope": "When candidate detectors or configurations are uncertain, select, tune, or combine candidate approaches to obtain more reliable anomaly judgments."
    },
    {
      "direction_id": "G11",
      "name": "Anomaly detection evaluation and assessment",
      "scope": "Use benchmarks, metrics, or systematic comparisons to assess anomaly detectors’ capabilities, limitations, and evaluation conclusions, rather than directly selecting a detector for a particular dataset."
    }
  ]
}