Model Capability Matrix¶
Use this table to choose a metric profile and evaluation path for each preset.
Source of truth
Run python -m seahorse evaluate metrics --help and
python -m seahorse evaluate surface --help for the exact controls in
your installed version.
Capability Flags¶
| Flag | Meaning |
|---|---|
| NLL | Per-event log-likelihood: Exact (closed-form or ODE-integrated compensator), Quadrature (density evaluated directly, compensator by numerical quadrature), Approximate (sample-based or surrogate value) |
| Sampling | Next-event predictive sampling (predict_next) |
| Surface | Intensity/density grid for evaluate surface |
Matrix¶
| Family | Python class | CLI preset | NLL | Sampling | Surface profile |
|---|---|---|---|---|---|
| AutoSTPP | AutoSTPP |
auto_stpp |
Exact | Yes | history_frame |
| DeepSTPP | DeepSTPP |
deep_stpp |
Exact | Yes | history_frame |
| NSMPP | NSMPP |
nsmpp |
Quadrature | Yes | — |
| SMASH | SMASH |
smash |
Approximate | Native | — |
| Diffusion STPP | DiffusionSTPP |
diffusion_stpp |
Approximate (sample KDE) | Native | — |
| Neural-STPP GMM (jump-ODE state, conditional-GMM decoder) | STPPEstimator("njsde") |
njsde (alias neural_stpp_gmm) |
Exact | Yes | future_exact |
| Neural JumpCNF | NeuralJumpCNF |
neural_jumpcnf |
Exact | Yes | future_exact |
| Neural AttnCNF | NeuralAttnCNF |
neural_attncnf |
Exact | Yes | future_exact |
| RMTPP + GMM | RMTPPGMM |
rmtpp_gmm |
Exact (factorized) | Yes | — |
| THP + GMM | THPGMM |
thp_gmm |
Quadrature (factorized) | Yes | — |
| Poisson + GMM | PoissonGMM |
poisson_gmm |
Exact (factorized) | Yes | — |
| Hawkes + GMM | HawkesGMM |
hawkes_gmm |
Exact (factorized) | Yes | — |
| Self-correcting + GMM | SelfCorrectingGMM |
selfcorrecting_gmm |
Exact (factorized) | Yes | — |
| Poisson + CNF | PoissonCNF |
poisson_cnf |
Exact (factorized) | Yes | — |
| Hawkes + CNF | HawkesCNF |
hawkes_cnf |
Exact (factorized) | Yes | — |
| Self-correcting + CNF | SelfCorrectingCNF |
selfcorrecting_cnf |
Exact (factorized) | Yes | — |
| Poisson + TVCNF | PoissonTVCNF |
poisson_tvcnf |
Exact (factorized) | Yes | — |
| Hawkes + TVCNF | HawkesTVCNF |
hawkes_tvcnf |
Exact (factorized) | Yes | — |
| Self-correcting + TVCNF | SelfCorrectingTVCNF |
selfcorrecting_tvcnf |
Exact (factorized) | Yes | — |
| Static Gaussian Hawkes | STPPEstimator("static_hawkes") |
static_hawkes |
Exact | Yes | — |
| Gaussian Poisson | STPPEstimator("gaussian_poisson") |
gaussian_poisson |
Exact | Yes | — |
| DNSK + Barrier (available; not benchmarked in the paper) | STPPEstimator("dnsk") |
dnsk |
Quadrature | Yes | — |
| GLIDE (available; not benchmarked in the paper) | STPPEstimator("glide") |
glide |
None (diffusion objective) | Native | — |
Choose This If…¶
| Goal | Recommended preset(s) |
|---|---|
| Smoke-test data and CLI wiring | poisson_gmm |
| Classical self-exciting baseline | hawkes_gmm |
| Flexible spatial density | *_cnf variants |
| Paper-style AutoSTPP reproduction | auto_stpp |
| Neural exact-density + surface diagnostics | njsde, neural_jumpcnf, neural_attncnf |
| Score-matching / generative experiments | smash, diffusion_stpp |
Compatibility Notes¶
njsdeis the canonical preset for the conditional-GMM neural exact family. Older aliases such asneural_cond_gmmare compatibility names — do not use them in new work.- SMASH and Diffusion STPP report approximate NLL (SMASH: framework surrogate; DSTPP: joint sample-KDE estimate, not a bound); note this in benchmark comparisons.
- NSMPP and THP + GMM evaluate their density directly but integrate the compensator by numerical quadrature (
nll_path="quadrature");EventCapabilities.nll_labelreports these labels. future_exactsurface profiles may prefer--device cpufor numerical stability on some hardware.python evaluate()exposes likelihood metrics only. Use the CLIevaluate metricscommand for full benchmark-aligned metric profiles.