Organizations: Nanyang Technological University, Singapore · Institute of Advanced Intelligence and Computing (IAIC), A*STAR, Singapore · Department of Computer Science, Khalifa University, UAE · Singapore University of Technology and Design, Singapore
Time-series classifiers deployed in risk-sensitive domains must be (1) accurate, (2) interpretable, and (3) able to abstain when evidence is weak. Existing methods do not deliver all 3 at once, and the obstacle is structural, rather than logistical. Meaningful temporal patterns cannot be defined without knowing segment boundaries, yet meaningful boundaries cannot be placed without knowing which patterns to expect. Existing approaches break this circularity by using fixed windows or expert-annotated vocabularies, thereby sacrificing at least one of the 3 properties. We introduce ConceptTime, which closes the loop through a single self-supervised signal. A frozen probabilistic forecaster predicts a distribution over the next observation at every timestep of the sequence. We call the mismatch between its prediction and the realized signal the predictive surprise. This signal both locates segment boundaries and characterizes each resulting segment through the forecaster. Segments are thus grounded by how they behaved relative to expectation, not by what they look like. These summaries cluster into a vocabulary of dynamical regimes that we call concepts. A lightweight head classifies from the concept sequence, and the distance to the nearest concept yields reliability by design. Across 17 public benchmarks, ConceptTime is competitive with state-of-the-art black-box accuracy, beats every interpretable baseline, and retains near-98% accuracy on Epilepsy with 1% of labels. It further remains competitive with softmax, prototype-distance, and SHAP/TimeX/TimeX++ on OOD detection and deletion faithfulness, without a single annotation, language model call, or domain expert. Code is available at: https://github.com/Sachithx/ConceptTime.git.
Figures & tables
Figure 1: ConceptTime targets accuracy, interpretability, and calibrated abstention through a single self-supervised primitive. Left , existing methods emphasize different goals. Right , 4 segments (A-D) and their preceding context. Raw-shape space conflates A and B (both end calm) and splits C from D. Predictive-geometry space, spanned by predictive variance and NLL, collapses A, C, and D as expectation is satisfied and isolates B as a surprise . Concepts grounded in deviation from prediction, therefore, distinguish dynamical regimes that shape-based vocabularies cannot.
Figure 2: Overview of ConceptTime . Stage 1: Autoregressive Gaussian Density Model (ARGDM) Mθ , trained once on raw signals, produces predictive surprise st that drives length-constrained DP segmentation. Each segment is summarized by a signature vector that combines predictive moments , residual morphology , and the surprise trajectory . Then, signatures are clustered into M concept prototypes; textual labels are post hoc, not training signals. Stage 2: new inputs are segmented by Mθ , mapped to concepts by nearest-prototype assignment, and classified through the K×M concept tensor. Each segment’s distance to its nearest prototype yields the fidelity score used for abstention.
UCI-HAR
Epilepsy
Sleep-EDF
UEA (10 avg)
Group
Method
1%
5%
Full
1%
5%
Full
1%
5%
Full
1%
5%
Full
Black-box
ROCKET
82.21
89.65
93.91
93.68
92.25
97.15
70.21
73.67
76.38
55.69
59.87
72.30
MiniRocket
86.55
95.63
97.31
94.71
96.86
98.26
79.53
82.15
85.91
51.90
54.89
73.80
InceptionTime
86.38
91.03
95.32
95.14
97.28
98.21
77.64
81.11
84.62
48.37
50.80
67.42
TS-TCC
70.23
75.82
93.72
91.20
93.20
97.12
75.26
78.98
85.67
62.37
63.87
72.71
TimesNet
85.33
90.05
91.82
93.74
96.33
97.57
52.36
55.62
75.88
60.01
62.28
68.76
Table 1: Classification accuracy (%) under full, 1%, and 5% label budgets. Red : best overall; Yellow : second best overall; Blue : best interpretable method per column. Shaded rows are ours.
Setting
A → B
A → C
A → D
B → A
B → C
B → D
C → A
C → B
C → D
D → A
D → B
D → C
Mean
#1st
Panel A. Zero-shot and few-shot transfer
Direct, 0%
36.25
44.46
29.58
50.62
81.82
98.57
52.05
87.84
94.20
57.19
93.60
85.15
67.61
0
Source- Mθ , 5%
90.14
75.84
91.06
90.40
84.02
98.53
96.66
97.43
97.62
88.60
92.00
90.74
91.09
2
Target- Mθ , 5%
95.42
91.20
97.29
93.73
90.54
98.72
96.08
98.97
98.28
87.94
95.06
89.17
94.37
7
TSLANet, 5%
85.89
82.04
86.36
86.73
87.87
99.56
89.11
96.30
96.00
90.65
99.85
88.78
90.76
3
Panel B. Density model and vocabulary adaptation with 5% target labels
Table 2: Target test accuracy (%) across FD operating conditions. Within each panel, Red and Yellow indicate the highest and second-highest accuracy for each direction, respectively.
Figure 3: OOD detection on HAR under Gaussian-noise and temporal-shuffle perturbations.
Figure 4: Deletion-based faithfulness on Epilepsy and HAR. Patches are removed in ranked order; lower AUC indicates more faithful attribution.
Signature components
Vocabulary size ( M )
Boundary signal
Dataset
Dist.
Ent.
Morph.
Mom.
Full
Surp.f.
8
16
32
64
128
Ent.
KL
Surp.
UCI-HAR
72.8
61.2
42.7
87.7
82.8
85.7
80.2
84.4
85.7
86.3
86.8
83.4
81.6
82.5
Epilepsy
97.2
96.1
94.6
96.4
96.7
98.0
98.1
98.2
98.0
98.4
98.5
97.4
97.6
98.1
Sleep-EDF
72.3
65.4
59.0
74.4
78.3
74.8
73.7
75.0
74.0
75.4
76.2
73.8
75.0
74.5
Table 3: Unified ablation on UCI-HAR, Epilepsy, and Sleep-EDF (full-supervision accuracy, %). Within each ablation group, Red marks the best value per dataset and Yellow the second best.
Figure 5: Learned concepts, patch boundaries, and predictive signatures on Epilepsy.
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Property
Post-hoc
CBMs
Adaptive Patch
Prototype / Shapelet
IB / Repr. Learning
ConceptTime (Ours)
Decision routes through concepts
○
●
○
◐
○
●
No external semantic supervision required
●
◐
●
●
●
●
Concepts discovered from data, not predefined
◐
◐
–
◐
●
●
Adaptive segmentation
○
○
●
○
○
●
Local, segment-level explanations
◐
◐
○
●
○
●
Built-in abstention without auxiliary head
○
◐
○
○
◐
●
Appendix
Table 4: Comparison of major directions in interpretable time-series classification. ● = direct support; ○ = absent; ◐ = partial or with caveats.
Dataset
Domain / Modality
#Train
#Test
#Ch.
T
Task / Classes
UCI HAR [ 3 ]
Wearable inertial sensing
7352
2947
9
128
6 activities
UCI Epilepsy [ 2 ]
EEG
9200
2300
1
178
Seizure / non-seizure
Sleep-EDF [ 19 , 12 ]
Sleep EEG
25612
8910
1
3000
5 sleep stages
UEA subset [ 6 , 63 ]
Mixed multivariate time series
Varies
Varies
Varies
Varies
10 classification datasets
Fault Diagnosis [ 31 , 49 ]
Bearing fault diagnosis
8184
2728
1
5120
3 classes; 4 transfer domains
Appendix
Table 5: Summary of datasets used in the experiments.
Dataset
#Ch.
T
Split
Domain
EthanolConcentration
3
1751
261 / 0 / 263
Alcohol industry
FaceDetection
144
62
5890 / 0 / 3524
Face / sensor signals
Handwriting
3
152
150 / 0 / 850
Handwriting
Heartbeat
61
405
204 / 0 / 205
Heartbeat
JapaneseVowels
12
29
270 / 0 / 370
Voice
PEMS-SF
963
144
267 / 0 / 173
Transportation
Appendix
Table 6: UEA multivariate archive subset used in our evaluation. The split column follows the format #train/#val/#test before any validation split is carved from the training set.
Dataset
Labels
TSLANet
ConceptTime -Dynamic
UCI-HAR
Full
95.41 ± 0.63
88.91 ± 3.40
1%
86.31 ± 1.42
74.60 ± 1.76
5%
90.99 ± 1.08
79.60 ± 1.31
Epilepsy
Full
97.26 ± 0.51
98.39 ± 0.35
1%
95.41 ± 1.19
97.91 ± 0.96
5%
96.28 ± 0.87
98.01 ± 0.71
Appendix
Table 7: Test accuracy (%, mean ± standard deviation across three seeds) for TSLANet and ConceptTime -Dynamic.
Unlabeled fraction
1% labels
5% labels
Val. NLL ↓
1%
95.64 ± 0.16
96.74 ± 0.50
-1.422
5%
95.78 ± 1.81
97.12 ± 0.78
-1.631
10%
96.46 ± 0.93
97.42 ± 0.37
-1.839
25%
97.07 ± 0.60
97.22 ± 0.29
-1.970
50%
97.33 ± 0.41
97.19 ± 0.25
-2.223
100%
97.36 ± 0.35
97.30 ± 0.26
-2.233
Appendix
Table 8: Unlabeled-data scaling on Epilepsy. Classification entries give mean test accuracy (%) ± standard deviation across 3 seeds. The last column gives the density model’s validation negative log-likelihood (NLL). In each accuracy column, red marks the highest mean and yellow the second highest.
Dataset
Size
Layers
d
Parameters
Accuracy (%)
Epilepsy
Tiny
2
64
110,149
98.65 ± 0.47
Default
4
128
810,629
98.39 ± 0.35
Larger
6
128
1,204,357
97.22 ± 0.19
UCI-HAR
Tiny
2
64
108,597
89.20 ± 3.32
Default
4
128
807,413
88.91 ± 3.40
Larger
6
128
1,201,141
86.94 ± 2.06
Appendix
Table 9: Density model size and downstream test accuracy (mean ± standard deviation across 3 seeds). red marks the best value and yellow the second best.
UCI-HAR
Epilepsy
Method
1%
5%
1%
5%
CT-FullU
65.73 ± 0.45
68.86 ± 0.31
96.64 ± 0.41
96.96 ± 0.24
CT-LabelOnly
69.16 ± 2.44
67.33 ± 2.61
96.22 ± 0.46
97.36 ± 0.09
CT-DensityFull/GMM-LabelOnly
64.25 ± 1.27
67.14 ± 1.47
96.16 ± 0.27
97.10 ± 0.37
TS-TCC
46.85 ± 5.57
85.25 ± 1.25
81.43 ± 2.60
93.25 ± 0.40
TS-TCC-SSL
72.51 ± 4.46
80.55 ± 1.72
94.28 ± 0.93
95.30 ± 0.17
Appendix
Table 10: Matched-label test accuracy (%, mean ± standard deviation across seeds 2021–2023). Column headings give the fraction of training labels used.
Protocol
1%
5%
CT-FullU
74.63
79.88
CT-LabelOnly
75.25
76.09
CT-DensityFull/GMM-LabelOnly
77.13
79.91
Appendix
Table 11: UCI-HAR test accuracy (%) on the additional matched split (seed 2026). Column headings give the fraction of training labels used.
Source
Pairs
Direct
Source- Mθ
Gain
Target- Mθ
Gain
FD-A/B/C
9
63.93 ± 26.58
91.30 ± 7.52
+27.37
95.58 ± 3.16
+31.65
FD-D
3
78.65 ± 19.06
90.45 ± 1.72
+11.80
90.72 ± 3.81
+12.08
All
12
67.61 ± 24.98
91.09 ± 6.47
+23.48
94.37 ± 3.84
+26.76
Appendix
Table 12: FD target test accuracy (%) by source condition, aggregated from Panel A of Table 2 . Entries give mean ± sample standard deviation across directions. Gains are percentage points relative to direct transfer within each row. Both few-shot settings use the source vocabulary.
Score
AUROC ↑
AUPR-OOD ↑
FPR95 ↓
Det. acc. ↑
Fidelity
0.666 ± 0.187
0.674 ± 0.186
0.719 ± 0.343
0.605 ± 0.145
Mean density NLL
0.533 ± 0.291
0.583 ± 0.224
0.770 ± 0.351
0.562 ± 0.116
Max density NLL
0.513 ± 0.163
0.522 ± 0.122
0.887 ± 0.170
0.509 ± 0.030
1−maxyp(y∣x)
0.528 ± 0.146
0.561 ± 0.100
0.935 ± 0.109
0.546 ± 0.047
Predictive entropy
0.517 ± 0.152
0.560 ± 0.108
0.922 ± 0.122
0.546 ± 0.047
Mean predictive variance
0.522 ± 0.255
0.543 ± 0.160
0.787 ± 0.312
0.505 ± 0.019
Appendix
Table 13: Detection of FD operating-condition shifts. Entries give mean ± standard deviation across the 12 directed pairs. AUPR-OOD treats target examples as positive. FPR95 is the false-positive rate at 95% OOD recall.
Method
Confidence score
AURC ↓
E-AURC ↓
ConceptTime
Fidelity
0.01060 ± 0.00164
0.01032 ± 0.00165
Mean prototype proximity
0.01062 ± 0.00198
0.01034 ± 0.00198
Max softmax
0.00144 ± 0.00012
0.00117 ± 0.00012
TOTEM
Mean codebook proximity
0.01514 ± 0.00284
0.01319 ± 0.00289
Max codebook proximity
0.01597 ± 0.00381
0.01401 ± 0.00386
Max softmax
0.01547 ± 0.00252
0.01351 ± 0.00247
Appendix
Table 14: Selective prediction on Epilepsy. Lower values indicate better ranking of correct predictions ahead of errors. Entries give mean ± standard deviation across seeds 2021–2023. VQShape’s deterministic logistic classifier gave identical predictions across these seeds. red marks the best value and yellow the second best.
Figure 6: Seizure and non-seizure recordings in concept space.
Figure 7: Concept assignment confidence across segments of an Epilepsy recording.
Figure 8: Segment boundaries overlaid on Epilepsy recordings.
Figure 9: Examples of adaptive segments in Epilepsy recordings.
Figure 10: Signature features of the learned Epilepsy concepts.
Figure 11: Examples of adaptive segments in UCI-HAR recordings.
Figure 12: Concept assignments for walking and walking upstairs on UCI-HAR.
Figure 13: Abstention decisions (left) and fidelity-weighted segment-to-concept assignments (right).
Department of Electrical Engineering Stanford University Stanford, CA, USA · Department of Computer Science University of California, Davis Davis, CA, USA