TU Dortmund's SYRAN Turns Anomaly Detection Into Readable Math Equations
SYRAN swaps opaque neural anomaly detectors for ensembles of short, human-readable equations that stay constant on normal data.
ProTakeaways
- SYRAN detects anomalies by learning short equations that stay near 1 on normal data.
- Loss combines invariance, a noise-contrast margin, and a complexity penalty to avoid trivial constants.
- Rediscovers Kepler's third law from 13 orbital samples in about one minute.
- Matches state-of-the-art AUC-ROC on 19 ADBench datasets when using best ensemble members.
- Produces readable formulas that recover known medical indicators and expose weak benchmarks.
- Open source at github.com/KDD-OpenSource/SYRAN; limited to datasets with modest feature counts.
Most anomaly detection models tell you something is wrong but not why. The pattern of what counts as normal lives inside millions of neural network weights or a forest of decision trees, and any explanation you get is bolted on after the fact. A group from TU Dortmund is pushing in the other direction with SYRAN, a method that represents normality as a small set of algebraic equations you can read, check, and hand to a domain expert.
Anomalies as broken equations
The core idea borrows from physics. SYRAN learns a collection of scalar functions that are approximately constant on normal data and uses their deviations as anomaly scores. The authors call these functions symbolic invariants, echoing Emmy Noether's view of physical systems as things defined by conserved quantities. If f(x) should equal 1 for every normal sample, then any test point where f(x) drifts far from 1 is suspicious.
Prior work like DEAN learns such invariants as deep networks. SYRAN instead instantiates the surrogate as symbolic expressions obtained via symbolic regression. The search runs over expression trees built from operators like +, -, *, /, sin, cos, and exp, so the output is a closed-form equation rather than a weight matrix.
The trivial-solution trap
Unsupervised symbolic regression has an obvious failure mode: the constant function f(x) = 1 satisfies the invariance objective perfectly and detects nothing. Identities like sin(x)^2 + cos(x)^2 = 1 are equally useless. SYRAN's loss function is built around avoiding these degenerate solutions by combining three terms: