[Question] Bug? IntensityFree: get_dt_stats computes statistics on raw inter-event times instead of log(τ), breaking the LogNormalMixtureDistribution standardization

[freshjarman]

Summary

The statistics fed into IntensityFree's LogNormalMixtureDistribution are computed on raw inter-event times $\tau$, but the variable is named mean_log_inter_time / std_log_inter_time and is consumed by an AffineTransform that — per the original IFL-TPP paper and reference implementation — expects the mean and std of $\log\tau$. This silently breaks the intended standardization of the GMM's base space and diverges from the upstream implementation (Shchur et al., ICLR 2020).

Where the issue is

1. easy_tpp/preprocess/dataset.py::TPPDataset.get_dt_stats

for dts, marks in zip(self.time_delta_seqs, self.type_seqs):
    dts = np.array(dts[1:-1 if marks[-1] == -1 else None])
    ...
    y_bar = dts.mean()    # mean of raw τ
    s_2_y = dts.var()     # var of raw τ
    ...
return x_bar, (s_2_x ** 0.5), min_dt, max_dt

Per-chunk inputs y_bar, s_2_y are computed on raw dts, not on np.log(dts).

2. easy_tpp/runner/base_runner.py (around L43)

mean_log_inter_time, std_log_inter_time, min_dt, max_dt = (
    self._data_loader.train_loader().dataset.get_dt_stats())
runner_config.model_config.set("mean_log_inter_time", mean_log_inter_time)
runner_config.model_config.set("std_log_inter_time", std_log_inter_time)

The values are assigned to keys that semantically promise "log-space statistics", with no log transform in between.

Comparison with the original IFL-TPP repo

Reference: shchur/ifl-tpp (ICLR 2020 official).

dpp/data/dataset.py computes log-space statistics explicitly:

def get_inter_time_statistics(self):
    """Get the mean and std of log(inter_time)."""
    all_inter_times = torch.cat([seq.inter_times[:-1] for seq in self.sequences])
    mean_log_inter_time = all_inter_times.log().mean()
    std_log_inter_time = all_inter_times.log().std()
    return mean_log_inter_time, std_log_inter_time

dpp/models/log_norm_mix.py documents the intended modelling chain:

x ~ GaussianMixtureModel(locs, log_scales, log_weights)
y = std_log_inter_time * x + mean_log_inter_time     # <- expects log-space stats
z = exp(y)

The class structure (AffineTransform ∘ ExpTransform) in EasyTPP is the same, but the values supplied to it are computed on the wrong scale.

Is it a bug?

[gespyrop]

I also believe that this is a bug.

I was trying to train IntensityFree on a different dataset and the log likelihood was a very large negative number (nowhere near the log likelihood reported by the rest of the models).

I tried to compute the mean_log_inter_time and std_log_inter_time on the training set which resulted in different numbers.

As a temporary fix, I tried monkey-patching the correct values and the model started working properly.

model_runner.model.mean_log_inter_time = mean_log_inter_time
model_runner.model.std_log_inter_time = std_log_inter_time

[iLampard]

Thanks for the excellent, well-researched report — confirmed on both counts, and it's actually a double bug: besides the raw-vs-log scale mismatch you identified, the streaming mean/variance combination in get_dt_stats incremented the sample count before the merge formulas that need the pre-update count, so even the raw-space statistics were aggregated incorrectly.

Both are fixed on main (2f1325a): get_dt_stats now computes the mean/std of log(clip(dt, 1e-5)) directly (the clip matches inter_times.clamp(min=1e-5) in IntensityFree.loglike_loss), matching the reference shchur/ifl-tpp implementation. There's also a new notebook, notebooks/easytpp_intensityfree_log_stats.ipynb, that reproduces both bugs and quantifies the impact — including why the log_scales clamp at [-5, 3] makes it impossible for the model to compensate on datasets with large raw time scales, which explains @gespyrop's observation exactly.

The fix ships in easy-tpp 0.2.3 on PyPI; until you upgrade, the monkey-patch in the comment above is the right workaround. Note that IntensityFree checkpoints trained before the fix aren't comparable with ones trained after, since the statistics are baked into the distribution.