MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series

AI summary

Overview

Research area: Generative modeling for time series — specifically flow matching applied to incomplete (partially observed) multivariate time series.

Technical level: Advanced. The paper assumes familiarity with flow matching, conditional flow matching objectives, diffusion models, Gaussian processes, and ODE-based sampling.

Scope: The paper proposes MissPath-FM, a flow-matching framework that redesigns the source distribution and the probability path so that generative transport concentrates on genuinely missing entries rather than re-transporting already-observed ones, and evaluates it across 48 experimental settings against eight baselines.

What This Paper Is About

Standard conditional flow matching starts from an uninformed Gaussian source, z_0 ~ N(0, I), and learns to transport every entry of a sequence from noise to the target. For partially observed time series this is wasteful: entries that are already known still have to be transported from noise, so the model spends capacity on regions where the answer is given. The paper's goal is to redesign two things — where the flow starts (the source prior) and how it moves (the probability path) — so that the velocity field concentrates on uncertain, missing entries while observed entries are locked in early.

Key Contributions

1. MissPath-FM framework. A flow-matching framework for partial observability that combines structured source priors (interpolation priors and Gaussian process posterior priors), mask and time-gap context conditioning, and a missingness-aware probability path.

2. A regime-dependent design hierarchy. The paper establishes that no single component suffices: the source prior yields a large initial gain under random missingness (MCAR) but can degrade performance under structured block missingness, where the learned flow becomes the primary driver; the missingness-aware path contributes a further 12%–26% across all regimes. Only the combination achieves robust performance.

3. A transport-geometric explanation. A formal transport cost decomposition (Proposition 1) plus trajectory measurements showing that structured priors shorten observed-region transport by up to 15 times and reallocate the velocity field's capacity to uncertain entries.

4. State-of-the-art performance across 32 settings at T = 48 (4 datasets, 4 missingness mechanisms, 2 ratios), winning 29 against eight baselines — 45/48 (93.8%) including T ∈ {24, 96}.

Main Findings

• Overall win rate: MissPath-FM wins 29 of 32 settings at T = 48 with an average rank of 1.16 across all methods, and 45 of 48 settings (93.8%) overall. The three losses occur at Electricity Block 30% and 50% (to BRITS and SAITS) and PhysioNet MNAR 50% (to SAITS, by 4%).

• Consistency across missingness types: It wins all 10 MCAR-type settings (including PhysioNet's artificial holdout), 6/8 Block, 7/8 MNAR, and 6/6 Irregular. Margins are largest under structured missingness: ETTm1 Block 50% (−27%) and Electricity MNAR 30% (−34%).

• Longer sequences help: At T ∈ {24, 96} on ETTm1, MissPath-FM wins all 16 settings, with margins growing at longer sequences (19% at T = 24, 26% at T = 96). On PhysioNet it wins 7 of 8 settings with narrow but consistent margins over SAITS.

• Trajectory-length ablation reveals a regime-dependent hierarchy: Under MCAR, the prior replacement gives the dominant gain (43%–85% MAE reduction), and the flow adds 9%–35%; under Block missingness the pattern reverses — interpolation-only fails on Electricity (MAE = 0.842, worse than Gaussian+Flow at 0.414) and is near-random on Synthetic (MAE = 0.857) — and the flow provides the dominant recovery (−67% on Electricity Block, −35% on Synthetic Block). Summarized: under MCAR the prior dominates on average (−65%) and flow is secondary (−21%); under block, flow becomes primary (−37%). The MA path contributes 18%–19% in both regimes.

• Every component is necessary: The incremental ablation shows each added component strictly improves MAE in all 8 reported rows. A leave-one-out ablation confirms this: removing the prior raises MAE by 373% on average, context encoding by 23%, the MA path by 20%, and RandomRate by 18%.

• The missingness-aware path contributes 12%–26% across all regimes, the one component whose benefit is consistent regardless of missingness pattern.

• Forecasting: At h = 24, MissPath-FM wins all 6 settings (three datasets, clean and 30% MCAR histories), with margins of 25%–35% on ETTm1 and Electricity. The Synthetic MCAR margin is narrow (0.572 vs 0.573), consistent with TSFlow's strength on smooth signals.

• Generation: MissPath-FM wins 8 of 9 metric–dataset combinations, achieving the best discriminative score on all three datasets (0.064–0.169). The single loss is the predictive score on Synthetic (0.261 vs Gauss+Std 0.240).

• Mechanistic evidence: On Synthetic Block 50%, the miss-to-obs trajectory-length ratio rises from 1.1× (Gaussian prior) to 19.2× (interpolation prior); observed entries travel roughly 15× shorter. The GP prior additionally shortens missing-region transport by 39%. Measured before the deterministic restoration step, observed-entry MAE is 0.023–0.025 for MissPath-FM versus 0.40–0.77 for a Gaussian-source flow, so observed values are held in place by the transport itself rather than by the final projection.

• Calibration and stability: Probabilistic calibration confirms 29/32 wins with ACE ≤ 0.104, and multi-seed evaluation shows CV < 2.4%.

• The Gaussian-prior inefficiency is formalized: Under a Gaussian source, each observed entry contributes μ² + σ² + 1, which equals 2 for z-scored data, so 𝒞_obs = 2·n_obs — growing linearly with the number of observed entries. At 50% missingness, roughly half the network's capacity is spent on trivially known entries. The interpolation prior reduces this to σ_obs²·n_obs → 0.

• The GP prior is an optional, costly alternative: All main results use the interpolation prior. The GP posterior prior is 23–39 times slower to construct, so the paper recommends it only for short, smooth series.

• Ablation control: A separate study retrains baselines with the same RandomRate augmentation and shows the augmentation does not help them, so the reported margins do not stem from it.

Methodology in Plain English

The paper begins with an analysis of where standard flow matching wastes effort. Under a Gaussian source, the expected squared transport distance splits into an observed part and a missing part, and the observed part scales with the number of observed entries — meaning the model spends effort moving values it already knows.

Three design changes address this:

1. A structured source prior. Instead of starting from pure noise, the flow is initialized close to the observed signal. The default is piecewise linear interpolation: observed entries keep their values, missing entries are interpolated between flanking observations (with forward/backward fill at boundaries). Noise is added with a small scale on observed entries (σ_obs = 0.1) and a large scale on missing ones (σ_miss = 1.0). An optional alternative replaces interpolation with a Gaussian process posterior over the full time grid, which provides joint temporal coherence and structured uncertainty that grows with distance from observed anchors.

2. A missingness-aware probability path. Rather than treating all entries identically along the flow, each entry gets its own transport speed: s_obs = 2 for observed entries and s_miss = 1 for missing ones. With these settings, observed entries complete 51% of their transport by t = 0.3 while missing entries complete 30%, and for t > 0.5 the observed residual is negligible. The path stays valid within the CFM framework (α(0) = 0, α(1) = 1, monotonically increasing, C¹ for all s ≥ 1).

3. A context-conditioned velocity network. A Transformer encoder runs along the temporal dimension; each token concatenates the interpolant, the mask, and a normalized time-gap tensor (3D dimensions total), projected to d_model and augmented with positional and flow-time embeddings.

Training minimizes a CFM loss over all entries plus a light auxiliary squared-error term on observed entries only (λ_obs = 0.1), which acts as a consistency safeguard rather than the mechanism that concentrates the velocity field. Inference integrates with fixed-step Euler (20 steps) and then deterministically restores observed entries.

Evaluation spans four datasets — Synthetic (D = 8, multivariate sinusoidal), ETTm1 (D = 7), Electricity (D = 20, top-variance UCI clients), and PhysioNet 2012 (D = 36, ~19% natural observation rate) — under four missingness mechanisms (MCAR, Block, MNAR, Irregular) at 30% and 50% ratios, with the default T = 48 and additional tests at T ∈ {24, 96} on ETTm1. All features are z-score normalized per variable and split 80/20 with fixed seed 42. A single global hyperparameter configuration is used for every dataset, mechanism, rate, and sequence length: a 4-layer Transformer encoder (d_model = 256, 4 heads, ~3.9M parameters), interpolation prior, MA path, λ_obs = 0.1, RandomRate training (r ~ U(0.1, 0.7) per sample), 500 epochs with AdamW, Euler-20, and K = 20 samples. Eight baselines are compared under identical splits and masks.

Why This Matters

Impact on research. The paper reframes missingness in generative time-series modeling: prior work largely treats it as an input-side conditioning signal (masks fed to a network, or a loss penalty on observed entries), whereas this work shows that the geometry of the transport itself is the bottleneck. The regime-dependent finding — that a source prior can actively hurt under block missingness while helping under MCAR — is a caution against assuming that better priors are uniformly better, and it gives a concrete, measurable diagnostic (trajectory arc length on observed versus missing entries).

Real-world applications.

• Clinical monitoring: PhysioNet 2012-style ICU data with ~19% natural observation rate and highly irregular per-variable sampling, where the model wins 7 of 8 settings and provides calibrated uncertainty alongside point predictions.
• Energy and grid telemetry: Electricity and ETTm1 settings, where sensor dropout is common and the largest margins appear under MNAR and block patterns (−34% and −27% respectively).
• Industrial sensor networks: Structured contiguous block gaps — the case where interpolation alone degenerates into flat extrapolation — are explicitly modeled, and the learned flow recovers performance there.
• Generative synthesis of sparse records: Unconditional generation results (best discriminative score on all three tested datasets, 0.064–0.169) suggest use in producing synthetic time series that preserve the statistics of sparsely observed real data.

Industry relevance. The method uses a single global configuration with no per-dataset, per-mechanism, or per-rate tuning, and requires no iterative denoising loop — inference is 20 Euler steps. That combination is attractive for deployment where missingness patterns vary across sensors, sites, or clients and per-case tuning is impractical. The paper also notes that a retrained-baseline control shows the RandomRate augmentation does not benefit the baselines, which matters for anyone comparing against these methods.

Future Directions

• Scaling. The largest tested setting is D = 36 and T = 192; behaviour at larger scale is explicitly not established.

• Extreme sparsity and noisy observations. The advantage shrinks when observations are very sparse or unreliable: against SAITS, MissPath-FM falls behind on two of four rate-sweep settings at 90% missingness, and under observation noise beyond about 0.2 (ETTm1) to 0.3 (Electricity) z-scored units.

• Cross-variable structure in the GP prior. The GP posterior prior is per-variable and ignores cross-variable dependencies — a natural extension would model joint structure across variables.

• Cheaper joint-coherence priors. The GP prior is 23–39 times slower to construct than interpolation, so it is recommended only for short, smooth series; a faster route to joint temporal coherence would broaden its applicability.

Target Audience

Researchers and practitioners in generative modeling for time series who already understand flow matching and diffusion. It is most useful for those working on imputation, forecasting, or synthesis under missing data, and for readers interested in how source distributions and probability paths — rather than network architecture or loss design — shape what a transport model learns. Readers without a background in continuous normalizing flows and ODE-based sampling will find the method sections dense, but the transport-cost argument and the ablation results are accessible at a conceptual level.

Authors’ abstract

Flow matching has emerged as a powerful alternative to diffusion models for generative modeling, but standard conditional flow matching is fundamentally misaligned with partially observed time series: starting from an uninformed Gaussian source forces the model to transport even already-observed entries from noise to the target, wasting capacity on trivially known regions. We propose MissPath-FM, a partial-observation-aware flow matching framework that redesigns the source prior and probability path for incomplete time series. MissPath-FM employs structured source priors (interpolation priors and Gaussian process posterior priors) to initialize the flow close to the observed signal, paired with a missingness-aware smooth path that allocates transport asymmetrically between observed and missing regions. We evaluate systematically across four datasets, four missingness mechanisms (MCAR, block, MNAR, irregular), two missing ratios, and three sequence lengths (T in {24, 48, 96}), totaling 48 settings against eight baselines. MissPath-FM wins 29 of 32 settings at T=48 and 45 of 48 overall (93.8%). Our ablations reveal a regime-dependent design hierarchy in which no single component suffices: under random missingness the source prior yields a large initial gain, but under structured block missingness it can fail entirely and the learned flow becomes indispensable; the missingness-aware path contributes a further 12%-26% across all regimes. Mechanistically, structured priors shorten observed-region transport by up to 15x, concentrating the velocity field on genuinely uncertain entries. Project page: https://zgp110.github.io/misspath-fm/