A frozen reconstructor becomes an imputer at test time: 19 of 20 dataset-backbone pairs cut error without retraining
Synopsis
When an entire sensor channel drops out, no local observations remain inside the window to interpolate from, and training a separate imputer for every failure pattern is impractical. The study treats each test window's missing entries as optimization variables, freezes the trained sequence reconstructor, and backpropagates the reconstruction residual on observed coordinates only, turning the reconstructor into an imputer at inference. Across five real datasets, four reconstruction architectures and one synthetically masked complete channel per test window, no-prior test-time optimization lowered mean absolute error in 19 of 20 dataset-backbone combinations, but beat mask-aware imputers outright on only two datasets.
Interpretation
A sequence reconstructor that has already been trained and frozen can be converted into an imputer at inference time for whole-channel missingness it never saw in training, without updating any weights. Deep imputers previously learned conditional completion or sampling during training, and mask-aware methods needed the observation mask plus artificial corruption training; a fully missing channel offers no within-window observations for local interpolation. On five real datasets and four reconstruction backbones, with one complete channel synthetically masked per test window, no-prior test-time optimization lowered mean absolute error relative to one-pass reconstruction in 19 of 20 combinations; the sole exception is TCN-AE on Electricity, where it rose 0.8%.
The missing positions become free variables, and the frozen model's reconstruction residual at observed positions is driven down to infer the missing values. The optimization variable starts from the training-set per-channel mean, observed entries are clamped back at every step, and gradients touch only missing coordinates, so the output is the model's completion rather than the raw optimized variable. Optimization uses Adam at a learning rate of 10⁻², at most 5000 iterations and objective-based early stopping; in the fastest configurations used for the runtime comparison, median stopping is 201 to 206 iterations and the 90th percentile is 201 to 209.
Whether the conversion works depends on missing inputs influencing observed outputs, which the study characterizes with the cross-Jacobian from missing inputs to observed residuals. For an identity-like reconstructor, observed outputs do not depend on missing inputs, the cross-Jacobian is zero, and the observed residual supplies no recovery gradient; weak or rank-deficient coupling leaves several completions equally self-consistent. When the cross-Jacobian has full column rank at an exact solution, that solution is isolated nearby and the local Gauss-Newton curvature is governed by the cross-Jacobian times its transpose; the optional temporal-smoothness prior only regularizes weakly constrained directions, lowering error in 11 of 20 combinations, leaving 2 unchanged at reported precision and raising 7.
Perspective
The demonstrated scope is five real datasets (BeijingAir, ItalyAir, ETTh1, Electricity, PeMS), four reconstruction backbones, and a controlled setting with one complete channel synthetically masked per test window; reconstructors train only on native missingness, while mask-aware baselines receive artificial corruption training.
The cross-Jacobian diagnostics cover three of the four backbones, with MTAD-GAT omitted because of its substantially higher diagnostic runtime; the authors describe these diagnostics as illustrative and do not establish that condition number alone predicts imputation accuracy or prior benefit. Evaluation does not cover multiple simultaneously missing channels, mixed block-plus-channel dropouts, or longer horizons, and does not establish performance for arbitrary reconstructors.
