Differentiable D-vine Copulas Enable Beam Search for Anomaly Detection
A new framework for fitting D-vine copulas uses gradient-based maximum likelihood estimation with a fully differentiable implementation, combined with beam search to maintain multiple competing configurations. This overcomes the combinatorial explosion of possible copula family and parameter choices as variables increase, avoiding the suboptimal local optima of sequential greedy methods. The approach is applied to localized anomaly detection, improving global fit by exploring diverse dependence patterns simultaneously.
Key facts
- Vine copulas model complex multivariate distributions via hierarchical bivariate pair-copulas
- Fitting D-vines requires selecting copula family and parameters from candidate sets
- Number of configurations grows combinatorially with variables and families
- Existing methods use sequential greedy decisions, potentially missing better global fits
- Proposed framework uses gradient-based MLE enabled by differentiable implementation
- Beam-search strategy maintains multiple competing D-vine configurations
- Applied to localized anomaly detection
- arXiv:2607.25020v1
Entities
Institutions
- arXiv