Troubleshooting Patterns¶

Problem	Solution
Data contamination or insufficient normal periods during training.	- Models tolerate small contamination but perform best with known-good (healthy) data. - Identify “normal” periods for training that cover all expected behavior variations. - See unsupervised models documentation.

Problem	Solution
Insufficient labeled examples, insufficient label specificity, or incomplete coverage of operating conditions.	- Provide enough labeled examples of both normal and abnormal behavior. - Ensure labels are accurate and do not overflow into unintended data patterns. - Ensure labels cover the full range of expected operating conditions. - Include events longer than the model’s window size to capture extended anomalies. - See supervised models documentation.

Problem	Solution
Sliding window parameter is too small to cover enough data points.	- Increase the window size. - Select a learning period without significant gaps or remove non-critical signals with gaps. - Ensure the window spans the smallest event of interest (≥ 10 sample points).

Problem	Solution
Pattern was not present in the training dataset; sensitivity depends on the generalization factor.	- Include the missing pattern in the training dataset. - Adjust the training period or generalization parameter. - For fully classified outputs: - Move generalization closer to 1.0 for broad matches. - Move closer to 0.1 for strict matches.

Problem	Solution
- Event region missing from learning period. - Not enough examples. - Events overlap other behaviors.	- Include the event region in the learning period. - Provide sufficient, non-overlapping examples. - Create separate event groups for each label. - Make labels specific to the intended behavior, avoiding overflow into other behaviors.

Problem	Solution
Model generalization settings or training data are misconfigured.	- Semi-Supervised Models: - Adjust number of examples and training periods. - Ensure labels capture unique behaviors only. - Unsupervised Models: - Stick to default cluster parameters unless signal behavior is well-understood. - Adjust generalization factor between 0.3–0.5 to control sensitivity.

Problem	Solution
Difficulty validating outputs across long time ranges or parameter choices.	- Compare outputs across parameter choices using SVD charts in Falkonry Reports. - Break long evaluation periods (> 6 months) into 1–2 month segments. - For supervised models, use the agreement score to measure alignment with the fact set. - Analyze condition-triggering distribution percentages; high distribution can increase false positives. See model evaluation documentation.