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

Yoo-Min Jung; Hyeon-Gi Kim; Jonghun Park

cs.LGSep 28, 2026

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

Authors: Yoo-Min Jung, Hyeon-Gi Kim, Jonghun Park

Organizations: Department of Industrial Engineering Seoul National University, Seoul, Republic of Korea

Abstract

Naturally irregular time series combine asynchronous observations, missing values, unequal lengths, and nonuniform sampling, while dense adapters can discard temporal structure. We propose a mask-aware state space classifier for irregular time series (MASCIT), which supplies observation masks to the encoder and excludes invalid steps from gated temporal aggregation. Across 34 irregular time series datasets, MASCIT yielded the strongest aggregate point estimate and was the only evaluated neural model with three-seed results on every dataset. MASCIT retained the lowest point rank across six overlapping irregularity indicators, while factorial ablations favored partial over full selectivity. These results support selective state space models as effective, executable backbones for naturally irregular time series classification.

Figures & tables

Figure 1 : Six irregularity indicators used in this study.

type	health			human activity recognition													mobility									sensor			other					synthetic
dataset	MI3	P12	P19	CT	GM1	GM2	GM3	GP1	GP2	GX	GY	GZ	LPA	PAM	PGZ	SGZ	AN	AOC	APT	ARC	GSS	MP	SE	TA	VE	DD	DG	DW	IW	JV	PGE	PL	SAD	ABF
US	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✓	✓	✗	✗	✓	✗	✗	✗	✓	✗	✓	✓	✓	✗	✗	✗	✗	✗	✓	✗	✗	✓
PO	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✓	✗	✗	✗	✗	✗	✗	✗	✓	✗	✗	✗	✓	✓	✓	✗	✗	✗	✗	✗	✗
VL	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✗	✓	✓	✓	✗	✗	✗	✓	✓	✓	✓	✓	✗
UL	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
SH	✓	✓	✓	✗	✗	✗	✗	✗	✗	✗	✗	✗	✓	✓	✗	✗	✗	✗	✗	✗	✓	✗	✓	✓	✗	✗	✗	✗	✗	✗	✓	✗	✗	✗

Table 1 : Dataset taxonomy based on six irregularity indicators for the 34 datasets used in this study.

\mathrm{UL}_{0}

denotes the original unequal-length indicator reported by PYRREGULAR and is retained for comparison with the refined taxonomy.

Figure 2 : Overall structure of MASCIT. MASCIT applies masks to the encoder input and temporal aggregation.

	non-deep learning						deep learning
	distance/kernel		tree	transform + tree			recurrent		neural CDE	graph	transformer	convolution	SSM
	KNN ( 1978 )	SVM ( 2016 )	LGBM ( 2017 )	ROCKET ( 2020 )	BORF ( 2024 )	RIFC ( 2026 )	BRITS ( 2018 )	GRU-D ( 2018 )	NCDE ( 2020 )	Raindrop ( 2022 )	SAITS ( 2023 )	TimesNet ( 2023 )	MASCIT (ours)
avg. F1 (%)	59.405	26.880	60.562	66.890	62.505	57.099	51.218	46.246	42.052	55.232	55.008	55.769	69.200
avg. rank	7.309	11.353	5.750	4.118	5.574	6.338	7.618	8.574	9.441	7.294	7.574	6.941	3.118
top-1	2	0	4	8	2	3	0	0	2	0	0	0	13
top-3	8	2	9	24	13	4	7	3	2	0	4	4	23

Table 2 : Aggregate F1 (%) performance and execution coverage across 34 irregular time series classification datasets. Bold : best, underline : second-best.

Figure 3 : Runtime versus average rank on the 29 datasets evaluated by all models; marker area shows data coverage.

Figure 4 : Dataset-level relative length dispersion versus the unadjusted generalist–irregular-neural macro-F1 gap (%p) on 29 common datasets.

Panel of Figure 5 — Figure 5 : Average ranks on 29 common datasets; the $\Delta$ in (a) reports the lower-row minus upper-row difference.

model	concat mask input	masked pooling	avg. F1 (%)	avg. rank	W/L	Wilcoxon $p$
MASCIT	✓	✓	69.200	1.71	–	–
	✗	✓	68.308	2.81	26/8	.022
	✓	✗	66.686	2.62	25/9	.006
	✗	✗	67.060	2.87	27/7	.005

Table 3 : Model ablation for the two mask interfaces. Bold : best, underline : second-best.

Figure 6 : Capacity ablation. Each cell value is an average F1 score (%). Red outlines $128\times 16$ .

Figure 7 : Configuration sensitivity versus sample size.

component	share	95% interval
$d_{\mathrm{model}}$	.411	[.336, .490]
$d_{\mathrm{state}}$	.013	[.008, .018]
time-(in)variance of $\Delta$	.228	[.155, .306]
time-(in)variance of $B$	.036	[.020, .057]
time-(in)variance of $C$	.015	[.008, .023]

Table 4 : Noise-adjusted functional ANOVA shares.

Figure 8 : Dataset-level TV-minus-TI effects for (top)

\Delta

, (middle)

B

, and (bottom)

C

. Only datasets with consistent directions across runs are shown.

Explore similar work

A Time-Aware Bag-of-Receptive-Fields for Interpretable Irregular Time Series Classification
Sep 30, 2026Francesco SpinnatoTime-Series Classification Time Series
Department of Computer Science, University of Pisa, Italy · ISTI-CNR, Pisa, Italy
BITS: Rethinking Fair and Comprehensive Evaluation for Irregular Time Series Forecasting
Sep 27, 2026Kangjia Yan, Linfeng Wang, Tianen Shen +7Time Series Time Series Foundation Models
East China Normal University, Shanghai, China. · University of Electronic Science and Technology of China, Chengdu, China. · Aalborg University, Aalborg, Denmark.
A Hybrid Attention Model Learning Unified Time-aware Patch Representation for Irregular Multivariate Time Series Forecasting
Sep 19, 2026Li Lin, Zhihao Lin, Qi Zhang +3Time Series Foundation Models
Southeast University, China · Nanyang Technological University, Singapore · Timecho Ltd., China

Sep 30, 2026, cs.LGJ/K move · Enter open · S save

A Time-Aware Bag-of-Receptive-Fields for Interpretable Irregular Time Series Classification

Francesco Spinnato

Department of Computer Science, University of Pisa, Italy · ISTI-CNR, Pisa, Italy

Time-Series Classification Time Series Time-Frequency Receptive Fields Bidirectional Long Short-Term Memory

Open paper AskPDF arXiv

Irregular time series, characterized by non-uniform sampling intervals, missing observations, and variable lengths, are ubiquitous in healthcare, mobility, and environmental monitoring, yet effective and interpretable classifiers for this setting are limited. Existing approaches often rely on imputation, which can obscure the temporal structure of the data, or require complex neural architectures that are opaque and difficult to explain. In this work, we extend the Bag-Of-Receptive-Fields (BORF), a fast, deterministic, and interpretable transform for time series, to the irregular setting. Our key contribution is a time-weighted normalization scheme in which each observation is weighted proportionally to its associated time delta, making pattern extraction sensitive to the actual temporal distribution of samples rather than only their index position. This requires deriving an efficient sliding-window recurrence for the time-weighted standard deviation, preserving the linear time complexity of BORF. We benchmark the resulting method against state-of-the-art irregular time series classifiers on datasets from the PYRREGULAR repository, demonstrating competitive classification performance with the added benefit of human-interpretable explanations.

Abstract

Explore similar work

A Time-Aware Bag-of-Receptive-Fields for Interpretable Irregular Time Series Classification

BITS: Rethinking Fair and Comprehensive Evaluation for Irregular Time Series Forecasting

A Hybrid Attention Model Learning Unified Time-aware Patch Representation for Irregular Multivariate Time Series Forecasting