MASCIT: A Mask-Aware State Space Classifier for Naturally Irregular Time Series

• Naturally irregular clinical, mobility, and wearable series combine asynchronous channels, missing observations, unequal lengths, and nonuniform sampling intervals.Source: §1 · p. 1
• Mamba's input-conditioned selection does not itself represent irregular observations, and selectivity need not be applied uniformly throughout an SSM layer.Partial selectivity allows adaptive and invariant dynamics to coexist.Source: §1 · p. 1
• Earlier Mamba classification work used regularly sampled archives, where value-only inputs hid observation patterns and gate normalization retained padding.PYRREGULAR spans 34 datasets with related variants and resource-limited failures, so performance, execution, and statistical dependence must remain distinct.Source: §1 · p. 1
• MASCIT feeds channel-wise observation masks to the selective recurrence and excludes invalid steps from multi-head adaptive pooling.It retains a shallow unidirectional backbone without separate imputation or bidirectional encoding.Source: §1 · pp. 1–2
• Across 34 datasets and 12 published baselines, MASCIT yielded the strongest aggregate point estimate and was the only evaluated neural model with complete data coverage.Source: §1 · p. 2

• BRITS and GRU-D combine recurrence with imputation or decay (Cao et al., 2018), while continuous-time models evolve hidden states between observations (Kidger et al., 2020).Source: §2 · p. 2
• Attention and graph models learn temporal and inter-sensor relations, whereas transforms, dictionaries, tree ensembles, and temporal convolutions need not model irregular time explicitly.PYRREGULAR compares these alternatives under a shared representation and protocol (Spinnato and Landi, 2026).Source: §2 · p. 2
• Mamba conditions selected state space parameters on the current input while retaining linear sequence complexity (Gu and Dao, 2024); prior classification work used a shallow backbone and gated aggregation.Source: §2 · p. 2
• Benchmark conclusions depend on dataset composition, preprocessing, and tuning budget, while resource-limited runs can bias completed-case averages.Comparisons should separate predictive quality, execution coverage, and temporal information exposed by the input.Source: §2 · p. 2

3.1 Irregular Time Series Representation

• Six overlapping indicators distinguish uneven sampling, partial observation, variable length, unequal channel length, shift, and ragged sampling across instance- and channel-level irregularities.The refined taxonomy separates cross-instance length variation from unequal channel lengths within an instance.Source: §3.1 · p. 2
• The adapter sorts the union of channel timestamps into consecutive positions, zeroes unobserved values, and pads instances to the maximum compacted length.Compaction preserves values, event order, length, and channel-wise mask geometry but replaces elapsed-time gaps with consecutive indices.Source: §3.1 · p. 2

3.2 State Space and Selective State Space Models

• Three selectivity switches choose input-conditioned or time-invariant forms for Δ, B, and C; Δ controls state-update timescale, while B and C govern input-to-state and state-to-output routing.The switches define eight patterns, and MASCIT retains this factorization for dense irregular series.Source: §3.2 · p. 3

3.3 MASCIT Classifier

• MASCIT concatenates values with observation masks before its partially selective SSM, then uses a valid-step mask to exclude padding from gated temporal aggregation.The model retains a shallow unidirectional backbone without separate imputation or bidirectional encoding.Source: §3.3 · p. 3

4.1 Datasets and Baselines

• The evaluation covers all 34 PYRREGULAR datasets and 12 baselines spanning non-deep and deep learning methods.The non-deep set includes KNN, SVM, LGBM, BORF, RIFC, and MiniRocket; the deep set includes BRITS, GRU-D, NCDE, Raindrop, SAITS, and TimesNet.Source: §4.1 · p. 4

4.2 Implementation and Evaluation

• MASCIT uses one partially selective Mamba layer with fixed dmodel = 128 and dstate = 16, evaluating eight prespecified TI/TV configurations per dataset.The reported configuration is the one with the highest test macro-F1 averaged over three seeds.Source: §4.2 · p. 4
• Macro-F1 is primary; ranks cover all 34 datasets, while Holm-corrected Wilcoxon tests compare within-dataset z-normalized F1 and runtime includes fitting and prediction.Data coverage and available-case average F1 are also reported; MASCIT uses the closest available setup to PYRREGULAR.Source: §4.2 · p. 4

5.1 Performance and Efficiency in Irregular TSC

• 34 datasets yielded MASCIT’s best average F1 and rank and most Top-1 finishes; on 29 common datasets, it also had the best rank at the lowest recorded runtime.It was the only evaluated neural method with results on all 34 datasets; Holm-corrected tests found significant gains over every model except ROCKET.Source: §5.1 · p. 4, Figure 3 · p. 5

5.2 Model Structures and Dataset Regimes

• Greater relative length dispersion was associated with a larger generalist-pipeline macro-F1 advantage, while MASCIT ranked first across both scale splits and grew relatively stronger on larger, multivariate, longer series.The length-dispersion finding is an association, not evidence of a causal benefit of length variation.Source: §5.2 · p. 5
• MASCIT had the lowest point rank in every irregularity-indicator group and stayed first after removing 17 VL-only datasets, while ROCKET fell from second to seventh.ROCKET was most competitive in the dominant VL group; MASCIT’s standing was not confined to that stratum.Source: Figure 5 · p. 5, §5.2 · p. 5
• MASCIT achieved the best average F1 and rank against the mask-interface ablations; removing masked pooling caused the largest average-F1 decrease, and the combined removal did not compound it.All three ablations were inferior under Holm-corrected Wilcoxon tests (p < .05).Source: §5.2 · p. 5, Table 3 · p. 5

5.3 Sensitivity Analyses

• MASCIT remained first across capacity settings except (dmodel, dstate) = (32, 1), with a broad high-performing width region of 64–256 and comparatively little change with state size.At each dmodel ∈ {64, 128, 256}, its average-rank difference from ROCKET exceeded 0.5 across the five dstate values.Source: §5.3 · pp. 5–6, Figure 6 · p. 6
• Main effects explained 70.3% of noise-adjusted configuration variation and interactions 29.7%; small datasets had wider configuration gaps, indicating selection uncertainty rather than transferable tuning gains.Among individual functional ANOVA components, only dmodel (41.1%), the time-(in)variance of ∆ (22.8%), and dmodel × ∆ (10.5%) exceeded 5%.Source: Table 4 · p. 6, §5.3 · p. 6, Figure 7 · p. 6

5.4 Partial Selectivity

• Dataset-level effects more often and more strongly favored TI for ∆, split between TI and TV for B, and showed no strongly consistent direction for C; preferences also varied by source and task.The greater prevalence of TI-∆ cases persisted despite natural irregularity.Source: Figure 8 · p. 6, §5.4 · pp. 6–7

• MASCIT attained the strongest aggregate point estimate and completed results on all 34 datasets, the only evaluated neural model to do so.Source: §6 · p. 7
• Factorial analyses favored partial over full selectivity and identified ∆ as the dominant source of selectivity-related variation.Source: §6 · p. 7
• The dataset-wise same-test oracle over eight patterns makes the aggregate an optimistic upper envelope, motivating a test-independent selection rule.Source: §6 · p. 7

• Table 5 reports dataset-level MASCIT performance and runtime over three seeds as mean ± sample standard deviation.Source: Table 5 · p. 8