acceptodds
Under review as a conference paper at ICLR 2027

CT-Miner: Cartesian Tree Pattern Mining for Time Series

Abstract

Time series often contain recurring structural patterns, and efficiently mining such patterns into compact representations is essential for scalable analysis of long sequences. Cartesian tree (CT) equivalence provides a well-established structural abstraction that preserves hierarchical order structure while discarding exact values and fine-grained ordinal variations. By grouping multiple ordinal patterns into a shared structural form, CT equivalence offers a principled way to compress recurring temporal structure. However, mining frequent CT-equivalent patterns at scale remains computationally expensive. A naive pairwise approach repeatedly constructs and counts CT representations over subsequences, requiring time for a sequence of length , which severely limits its applicability to long sequences. We propose a new Cartesian pattern mining algorithm based on a Cartesian suffix tree that compactly organizes CT-equivalent subsequences and reuses shared structural information. Our method reduces exhaustive CT-pattern occurrence collection from to time, and we formally prove the correctness and complexity bounds. We further show that this computational gain translates into effective compact representations. Across diverse time-series datasets, a small set of mined CT patterns preserves meaningful clustering structure, and comparisons with finer-grained order-preserving representations show that CT equivalence reduces redundant ordinal distinctions under limited feature budgets. Our implementation is available at https://anonymous.4open.science/r/CT-Miner-Submission-053B/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Pattern Mining, Cartesian Tree Equivalence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_VxZNOpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies frequent time-series pattern mining under Cartesian-tree (CT) equivalence, which merges some order-preserving patterns while retaining their hierarchical minimum structure. Using an existing Cartesian suffix tree, CT-Miner obtains occurrence counts from subtree leaf counts and represents patterns through compact node references. The paper claims O(n2)O(n^2) occurrence collection across all lengths after O(nlog⁡n)O(n\log n) index construction, and introduces Catalan-based integer labels for CT structures. Experiments construct top-KK occurrence-count features and evaluate k-means clustering on five main UCR datasets, with Wafer analyzed separately. Comparisons include order-preserving miners, PCA, and an independent-window CT-Hash implementation. The reported results show competitive clustering quality and reduced feature-construction time, together with substantial memory costs in some longer-pattern settings.

Critical strengths
  1. CT-Miner finds recurring patterns faster by reusing work across overlapping subsequences. In one comparison on Car, feature-construction time falls from about 26 seconds to 4 seconds.
  2. The method retains useful information with a small set of features. Using 30 patterns to describe each series, it achieves clustering quality comparable to existing pattern-mining methods.
  3. The features have a clear meaning: each records how often a particular local structure occurs. This makes it possible to inspect which structures differ between groups of time series.
Critical weaknesses
  1. The endmarker check in line 12 of Algorithm 1 does not match the stated indexing conventions. The condition used in the experiments needs clarification.
  2. CT-Hash rebuilds every window’s representation independently. Without an incremental counting baseline, the benefit of the CST over simpler reuse strategies remains unresolved.
  3. The OP-baseline timing comparison does not isolate the contribution of CT’s top-KK early stopping to the reported 6.6×6.6\times speedup.
  4. On PigCVP at L=256L=256, CT uses 6,841.0 MiB6{,}841.0\,\mathrm{MiB} versus 753.7 MiB753.7\,\mathrm{MiB} for CT-Hash. The paper acknowledges this cost but does not separate index, retained-pattern, and feature-construction memory.
Questions for authors
  1. Please clarify the correct endmarker check in Algorithm 1 and the condition used in the experiments.
  2. How does CT compare with exact incremental PD counting that reuses shorter-pattern IDs?
  3. How much of the 6.6×6.6\times speedup remains when CT’s top-KK early stopping is disabled?
  4. How much peak memory comes from the index, retained patterns, and feature construction?
4 comments

We thank the reviewer for identifying these issues. We corrected the endmarker convention and added incremental baseline, early stopping, and memory analyses.

W1. Endmarker convention

The reviewer is correct. Under our 1-based indexing convention and half-open interval notation, the endmarker condition in Algorithm 1 should be l+idx(n)=∣T∣+3l+\mathrm{idx}(n)=|T|+3, rather than ∣T∣|T|. We have corrected this indexing error and clarified the treatment of endmarkers in the definitions of the parent-distance representation and the associated lemmas and proofs. The experimental implementation already excludes endmarker-containing windows through valid-window bounds, so these corrections do not affect the reported results.

W2. Incremental counting baseline

We added CT-Hash-Incremental, which reuses shorter-pattern parent-distance IDs instead of reconstructing every window. We compared it with CT-Miner and CT-Hash at K=100 with identical support-bound early stopping. For each of 18 dataset-length conditions, five runs completed with matching ordered dictionaries, supports, features, and stage signatures.

DatasetMax LCT-MinerCT-HashCT-Hash-IncrementalHash / CTInc. / CT
Beef80.350.410.331.18x0.97x
160.500.670.591.35x1.19x
320.721.341.041.88x1.46x
641.041.991.431.91x1.37x
1281.062.101.421.99x1.34x
2561.011.921.471.89x1.45x
Car80.460.540.451.17x0.97x
160.651.050.791.61x1.21x
321.082.111.771.95x1.63x
641.875.203.162.78x1.69x
1282.8210.705.273.79x1.87x
2562.7510.945.483.98x1.99x
PigCVP82.032.001.730.99x0.85x
163.376.195.541.84x1.65x
324.178.927.422.14x1.78x
644.268.916.842.09x1.61x
1285.2410.278.301.96x1.58x
2568.2916.7913.172.03x1.59x

CT-Miner is fastest in 15 of the 18 conditions. At L=256, it is 1.45x, 1.99x, and 1.59x faster than incremental counting on Beef, Car, and PigCVP, respectively.

W3. Effect of early stopping

Our early stopping criterion is an exact pruning strategy rather than heuristic early termination. It stops only when no unvisited pattern can enter the current top-K dictionary, thereby preserving the selected patterns and their frequencies. We further verified that early stopping produces identical results by comparing the selected pattern dictionaries and feature matrices with early stopping enabled and disabled. All 30 paired comparisons across six datasets and five repetitions produced identical results. The table reports mean wall-clock times and the maximum pattern lengths explored with early stopping enabled and disabled, allowing pattern lengths up to the full sequence length.

DatasetEarly on (s)Early off (s)Off / OnMax Pattern Length (on / off)
Beef0.351.614.59x18 / 470
Car0.444.239.59x18 / 577
Elec.Dev.1.888.004.25x6 / 96
MoteStrain0.751.221.63x22 / 84
PigCVP1.4062.9544.89x10 / 2000
Wafer2.1411.755.49x14 / 152

W4. Memory breakdown

We further analyzed the memory usage of CT-Miner by measuring the peak Java process RSS and the sizes of its major data structures. We report the results at L=256 with early stopping disabled and K=30, averaged over five runs. All values are in MiB. Tracked arrays are calculated array sizes at the sampled RSS peak. Untracked heap and non-heap residual are the remaining memory not attributed to these arrays.

DatasetPeak RSSCST index arraysPrefix-ID registry arraysInput & feature arraysUntracked heapNon-heap residual
Beef692.991.7686.400.23278.74325.86
Car847.894.39192.000.56448.53202.40
PigCVP6820.7236.883072.004.843395.79311.20

On PigCVP, the prefix-ID registry arrays occupy 3,072.00 MiB, corresponding to approximately 45.0% of the peak RSS, whereas the compact index arrays occupy only 36.88 MiB. The final feature matrix requires only 0.07 MiB. The peak occurs during the feature-construction phase in all five runs.

We additionally performed garbage-collection diagnostics on all three datasets. Explicit garbage collection reclaimed 325.98 MiB on Beef, 414.72 MiB on Car, and 3,157.06 MiB on PigCVP. These results indicate that substantial heap usage comes from temporary or obsolete allocations, potentially including arrays replaced during prefix-ID registry expansion.

Overall, the high peak memory usage reflects both the retained prefix-ID registry and substantial allocations awaiting garbage collection, rather than the compact index or final feature representation.

Thank you for addressing some of my concerns. The ablation measures the effect of early stopping on CT-Miner, but it does not show how much of the reported 6.6x speedup over OPF remains when early stopping is disabled. I’ve updated my rating from 2 to 3.

We thank the reviewer for reconsidering our work. We would like to address the reviewer's remaining concern regarding the impact of early stopping on the reported speedup over OPF.

We agree that early stopping contributes substantially to the practical efficiency of CT-Miner. Importantly, our early stopping criterion is exact rather than heuristic. It skips unnecessary computation without changing the selected top-K patterns or their occurrence counts. We therefore use early stopping in our primary evaluation, since it does not change the mining results as confirmed by our ablation. Thus, a comparison without early stopping is less representative of CT-Miner's practical performance.

Nevertheless, to quantify its impact, we compare the existing results in Tables 6 and 16. Across the five main datasets at K=30, OPF requires 8.55 seconds on average, while CT-Miner with early stopping disabled requires 15.60 seconds. Thus, without early stopping, CT-Miner is approximately 1.83x slower than OPF on average.

Time Comparison Between CT-Miner and OPF

DatasetOPF (s)CT-ON (s)CT-OFF (s)OPF/CT-ONOPF/CT-OFF
Beef2.050.381.615.42x1.27x
Car15.050.544.2327.72x3.56x
ElectricDevices4.322.628.001.65x0.54x
MoteStrain3.540.991.223.56x2.90x
PigCVP17.781.9462.959.16x0.28x
Mean8.551.3015.606.60x0.55x

Thank you for your detailed response. My score is already positive.

Reviewer_sJoROpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a new Cartesian pattern mining algorithm based on a Cartesian suffix tree that compactly organizes CT-equivalent subsequences and reuses shared structural information, reducing the time complexity from O(n4)O(n^4) to O(n2)O(n^2).

Clustering experiments show that small CT pattern dictionaries retain meaningful structural information under limited feature budgets.

Critical strengths

This paper formally proves the correctness of the proposed mining procedure and these complexity bounds.

Critical weaknesses

The comparison is limited to OPP mining, making it difficult to assess the advantages over other time-series representation methods. Please compare performance with other representation methods on downstream tasks or demonstrate an application where OPP mining is particularly effective.

For the former, please compare classification accuracy and speed against existing methods on standard benchmark such as UCR/UAE. Examples of representative methods include local discriminative representations (e.g., shapelets), random convolutional features (e.g., MiniROCKET), and self-supervised representations (e.g., TS2Vec).

For the latter, please identify the tasks and data characteristics for which OPP mining is most effective, and compare its performance against other representation methods on these tasks and datasets.

Questions for authors

Please address the weaknesses.

7 comments

We thank the reviewer for requesting comparisons beyond ordinal pattern mining. We evaluated classification on six UCR datasets, forecasting on six datasets, and anomaly detection on 250 UCR-Anomaly cases.

W1. Classification and runtime

We compared CT against raw values, PCA, statistical and ordinal features, and shapelet-based and convolutional representations. The table gives mean classification accuracy, macro-F1, and end-to-end pipeline time over six datasets.

Classification

MethodACC (%)Macro-F1 (%)Time (s)
Raw85.785.70.002
PCA88.187.90.004
Stats56.054.70.002
CT67.465.90.385
CT + Stats69.067.40.385
OPP68.567.01.056
EFO68.567.00.905
RDST92.692.41.610
MiniROCKET93.092.90.190

CT achieves accuracy similar to OPP/EFO while reducing mean pipeline time by approximately 2.7x and 2.4x, respectively. However, MiniROCKET and RDST achieve much higher accuracy. Runtime comparisons reflect the evaluated implementations rather than hardware-normalized algorithmic speedups.

W2. Additional downstream tasks

We additionally evaluated forecasting and anomaly detection to further investigate the applicability of CT-based representations. Forecasting results averaged across six datasets and four prediction horizons further reveal the limitations of structural representations. CT + Stats achieves an average normalized MSE of 0.315, compared with 0.173 for Raw + Ridge and 0.297 for TimesFM 2.5. These results suggest that CT-based representations discard magnitude information important for accurate forecasting.

We also evaluated anomaly detection on 250 UCR-Anomaly cases, where CT and OPP-based representations demonstrate advantages in capturing structural anomalies. The table reports mean AUROC and AUPRC across 142 original, 92 distorted, and 16 noise cases.

Anomaly detection

MethodAUROC (%)AUPRC (%)
Raw + kNN82.028.1
Raw + IF61.613.1
Stats + kNN76.225.1
Stats + IF72.214.9
CT + kNN76.623.4
CT + IF73.115.7
CT + Stats + kNN84.531.7
CT + Stats + IF75.815.6
OPP/EFO + kNN76.022.9
OPP/EFO + IF72.515.0
MOMENT51.91.7
TS2Vec50.11.9

CT + Stats + kNN obtains the highest mean scores, improving on Raw + kNN by 2.5 AUROC and 3.6 AUPRC percentage points. On original cases its AUROC reaches 87.1% versus 82.3% for Raw, whereas on noise cases it falls to 75.7% versus 79.7%.

W3. Successful and unsuccessful cases

The following original cases illustrate both outcomes. Each entry gives AUROC / AUPRC in percent under the same kNN detector.

DatasetRawStatsCTCT+StatsOPP/EFO
WalkingAceleration16.0 / 2.749.5 / 4.7100.0 / 98.999.5 / 93.5100.0 / 98.9
InternalBleeding921.9 / 2.122.3 / 2.199.9 / 96.099.9 / 96.099.9 / 96.0
PowerDemand248.7 / 2.297.5 / 72.1100.0 / 99.5100.0 / 99.5100.0 / 99.5
apneaecg353.5 / 0.467.1 / 0.6100.0 / 97.3100.0 / 96.9100.0 / 97.7
2sddb4077.6 / 1.595.6 / 8.0100.0 / 95.299.9 / 90.999.0 / 28.3
3sddb4081.8 / 1.989.3 / 3.1100.0 / 95.999.9 / 90.899.2 / 29.6
gait372.2 / 2.980.0 / 4.295.3 / 35.0100.0 / 100.095.3 / 35.0
weallwalk-25099.4 / 10.99.1 / 0.10.2 / 0.11.9 / 0.151.1 / 0.2
Italianpowerdemand-21079.6 / 0.440.3 / 0.112.2 / 0.115.2 / 0.112.2 / 0.1
TkeepThirdMARS-160100.0 / 96.245.2 / 1.565.0 / 2.460.4 / 2.373.2 / 2.9
mit14134longtermecg-22896.6 / 1.888.1 / 0.327.4 / 0.045.6 / 0.148.0 / 0.1
CHARISten-20790.3 / 1.283.7 / 0.766.7 / 0.356.7 / 0.363.0 / 0.3

OPP-based representations perform well on selected anomaly detection datasets, sharing many favorable cases with CT. CT achieves comparable performance on these datasets while substantially improving AUPRC on 2sddb40 and 3sddb40. Moreover, CT + Stats outperforms Raw on average, achieving 84.5% AUROC and 31.7% AUPRC compared with 82.0% and 28.1%. Some cases, such as weallwalk, exhibit nearly inverted anomaly rankings, indicating strong discrimination with reversed score orientation. Other cases, including mit14134longtermecg and CHARISten, show limited discrimination even after score reversal. These results suggest that ordinal and hierarchical structural patterns can be effective for detecting certain anomalies, while the loss of magnitude information may limit their applicability to others.

Overall, CT-Miner offers efficient, training-free, and interpretable structural pattern mining, complementing rather than replacing learned time-series representations.

Thank you for the response. I appreciate the authors' detailed response. My understanding of your response is as follows. Please correct me if I am mistaken.

W1: With CT/OPP-based coarse-grained representations, ACC drops from 85.7 (raw data) to 69.0/68.5. Since MinROCKET achieves an ACC of 93.0, it appears less suitable for classification tasks.

W2: Accuracy decreases from 84.5 for CT+Stats+kNN to 76.0 for OPP/EFO+kNN, while runtime increases from 0.385 (CT+Stats) to 1.056(OPP), as shown in the response to W1. If your OPP improves time complexity over CT, showing its impact on run time would make the contribution more convincing.

W3: The strong dependence of accuracy on the dataset is particularly interesting. Clarifying which types of data and tasks CT/OPP is suitable for, and which it is not, would make the paper more useful to readers.

Regarding novelty, the impact of the proposed variants of CT/OPP on the broader ML community beyond the field of CT/OPP deserves consideration. However, I could not adequately assess this impact based on my expertise, so I did not factor it into my rating.

We thank the reviewer for the follow-up and for carefully considering our response. Your understanding is correct for W1 and W3 and there seems to be a confusion between our work (CT) and the previous work (OPP) regarding W2.

W1. Classification performance

As you noted correctly, CT-based representations do not provide competitive classification accuracy compared with raw inputs or methods such as MiniROCKET, which we acknowledge as a limitation and discuss in detail in Appendix D.10 (pp. 34–35).

W2. Runtime and anomaly detection

We would like to clarify that our method aims to efficiently mine CT-equivalent patterns, while OPP and EFO serve as comparison baselines. Our theoretical complexity improvement concerns CT-pattern mining, whereas the runtime advantage over OPP/EFO is demonstrated empirically. As you correctly observed, CT-based representations achieve better anomaly detection performance and lower runtime than OPP/EFO. The runtime values cited in your comment correspond to the classification experiments, so we additionally provide the anomaly detection runtimes below for completeness. The table reports mean end-to-end pipeline times across 250 UCR-Anomaly cases using the same kNN detector.

MethodAUROC (%)AUPRC (%)Time (s)
CT + Stats + kNN84.531.70.936
OPP + kNN76.022.92.740
EFO + kNN76.022.92.565

CT + Stats achieves 2.93x and 2.74x end-to-end speedups over OPP and EFO, respectively, while also improving anomaly detection performance. These results again demonstrate the practical benefits of CT-based pattern mining in both efficiency and downstream utility. We have added Appendix D.12 (pp. 36–39) with the anomaly detection evaluation and discussion.

W3. Dataset-dependent performance

We agree that the effectiveness of CT-based representations depends on the dataset characteristics. As discussed in our previous response, CT equivalence discards magnitude information, which can limit its effectiveness when numerical values are important. Following your suggestion, we have further analyzed successful and unsuccessful cases, including differences across anomaly types and representative datasets. These findings are discussed in Appendix D.12 (pp. 37–39), where we clarify the conditions under which CT-based representations are beneficial and their limitations.

Novelty and broader impact

We understand your position and appreciate to mention a deserved consideration of the broader relevance of CT-Miner to the ML community. Beyond the specific field of pattern mining, we believe CT-Miner has potential as an efficient, training-free auxiliary feature extractor that captures structural information complementary to conventional numerical features. This perspective is consistent with recent studies highlighting the utility of interpretable subsequence features ([1] Qu et al., 2024; [2] Liu et al., 2025), their complementary role in machine learning models ([3] Wen et al., 2025), and the continued practical relevance of simpler approaches to time-series anomaly detection ([4] Liu and Paparrizos, 2024). Overall, our results suggest that CT-Miner can contribute to broader time-series machine learning by providing efficient, complementary structural features beyond pattern mining.

[1] CNN Kernels Can Be the Best Shapelets, Qu et al., ICLR 2024

[2] Learning Soft Sparse Shapes for Efficient Time-Series Classification, Liu et al, ICML 2025

[3] Shedding Light on Time Series Classification using Interpretability Gated Networks, Wen et al., ICLR 2025

[4] The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark, Liu and Paparrizos, NeurIPS 2024

Reviewer_LsstOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CT-Miner, a framework for fast and coarse-grained time-series pattern mining based on Cartesian tree (CT) equivalence. The key idea is to represent time series substrings by their Cartesian tree structure, which captures hierarchical order relationships while discarding exact values and fine-grained ordinal variations. The authors introduce a Cartesian suffix tree (CST) data structure that enables linear-time construction and efficient counting of CT-equivalent pattern occurrences, avoiding the O(n^2 L) or O(n^4) costs of naive pairwise approaches. Experiments on five UCR-like time series datasets demonstrate that CT-based features achieve competitive clustering quality (NMI, homogeneity, ARI) against several frequent ordinal pattern baselines while requiring orders of magnitude less feature-construction time and memory.

Critical strengths

The primary strength is the elegant use of Cartesian suffix trees to achieve provably efficient CT pattern mining. By exploiting the shared prefix structure of suffixes, the CST avoids redundant Cartesian tree constructions across overlapping windows, reducing the worst-case complexity from O(n^2 L) to near-linear in the input size for practical pattern lengths. This is a clean algorithmic insight that transforms an otherwise intractable enumeration problem into a scalable pipeline. Building on this efficiency gain, the empirical evaluation is thorough and well-organized. The paper tests CT across five diverse datasets with varying sequence lengths and class counts, reports multiple clustering quality metrics, and provides detailed ablations on feature budgets and input prefix lengths. The consistency of CT's performance relative to strong ordinal-pattern baselines across all settings strengthens the claim that CT equivalence is a practically useful coarse-graining abstraction.

Critical weaknesses
  1. My major concern is the limited demonstration of CT's advantage in truly large-scale settings. All experiments use datasets with sequence lengths up to roughly 2,000 and at most ~17,000 series. The paper's theoretical motivation emphasizes scalability to long sequences, yet the CT-Hash baseline -- the natural naive comparator -- is only evaluated on three datasets (Table 3) and does not scale to the full benchmark suite. Without results on datasets with sequences exceeding tens of thousands of time steps, it remains unclear whether the CST-based approach maintains its computational advantage when hash collisions and memory pressure become significant, or whether alternative suffix-structure methods (e.g., suffix arrays adapted to ordinal patterns) might close the gap.

  2. Beyond the scale limitation, the clustering task used for evaluation is relatively simple and does not reflect the full complexity of downstream time series applications. Clustering is a proxy task; the paper does not evaluate CT on classification, forecasting, or anomaly detection -- tasks where representation quality is more critically tested. Moreover, the CT abstraction inherently discards magnitude information and fine-grained ordinal details. For applications where absolute values or subtle trend changes carry discriminative signal (e.g., financial time series or physiological signals), this information loss could be detrimental, yet the paper provides no analysis of when CT abstraction fails or degrades performance.

  3. In addition to these concerns, the comparison with recent time series representation learning methods is sparse. The paper benchmarks against classical ordinal pattern miners (OPP, EFO, OPF, OPR, SOPP) and PCA, which is appropriate for the pattern-mining niche. However, it does not compare against recent foundation-model-based or self-supervised representation learning approaches for time series (e.g., Time-MoE, or other transformer-based pretrained models) that have emerged as strong competitors in representation quality. Including at least a qualitative discussion of how CT features complement or compete with learned representations would strengthen the paper's positioning and clarify its intended niche.

Questions for authors
  1. How does the CST-based mining algorithm scale in practice when applied to time series datasets with sequence lengths exceeding 10,000? Could you report CT-Hash or naive baseline results on at least one such dataset to demonstrate the practical speedup of your approach?

  2. The CT abstraction discards magnitude information by construction. On which types of time series datasets (e.g., those with high intra-class magnitude variance) does this abstraction cause the most significant clustering quality degradation, and are there simple mitigations (e.g., augmenting CT features with magnitude statistics)?

  3. How do CT-based features compare with representations learned by recent time series foundation models or self-supervised methods? Even a qualitative discussion of complementary strengths would help position CT-Miner within the broader time series representation learning landscape.

2 comments

We extended scalability experiments and evaluated three downstream tasks against classical and learned baselines.

W1. Long-sequence scalability

We compared CT-Miner and CT-Hash on five UCR-Anomaly series at 10k and 20k and three at 50k. Five repetitions mined all lengths from 2 to L. Mean feature construction times are reported in seconds.

L10k CTHashSpeedup20k CTHashSpeedup50k CTHashSpeedup
80.330.310.94x0.440.410.92x0.530.470.88x
160.470.511.09x0.570.661.15x0.780.911.171x
320.620.951.52x0.841.361.61x1.292.361.83x
640.931.811.94x1.353.062.28x2.225.862.64x
1281.413.942.80x2.157.143.32x3.3915.554.59x
2562.2410.974.91x2.9718.506.23x5.7848.438.38x

CT-Hash is slightly faster at L=8, due to index overhead. CST reuse becomes increasingly beneficial as L grows. At L=256, CT-Miner gains 4.91x, 6.23x, and 8.38x at 10k, 20k, and 50k.

W2. Downstream applicability and limitations

We evaluated classification on six UCR datasets, forecasting on six datasets and anomaly detection on 250 UCR-Anomaly cases. Results are averages over datasets.

Classification

MethodACC (%)Macro-F1 (%)
Raw85.785.7
PCA88.187.9
Stats56.054.7
CT67.465.9
CT + Stats69.067.4
OPP/EFO68.567.0
RDST92.692.4
MiniROCKET93.092.9

Forecasting (Input length 96 with horizons 96, 192, 336, and 720)

MethodMSEMAE
Persistence0.3190.399
Raw + Ridge0.1730.304
Stats + Ridge0.2450.365
CT + Ridge1.2120.910
CT + Stats + Ridge0.3150.405
OPP/EFO + Ridge1.1940.919
TimesFM 2.50.2970.376
TimesFM 3.00.7540.484

Classification and forecasting expose CT's limitations. Its hierarchical order patterns discard magnitudes important for class discrimination and numerical prediction. Adding mean, standard deviation, minimum, and maximum helps but does not fully restore predictive information. Learned classifiers and raw-value regression remain substantially stronger.

Anomaly detection (142 original, 92 distorted, and 16 noise)

MethodAUROC (%)AUPRC (%)
Raw + kNN82.028.1
Raw + IF61.613.1
Stats + kNN76.225.1
Stats + IF72.214.9
CT + kNN76.623.4
CT + IF73.115.7
CT + Stats + kNN84.531.7
CT + Stats + IF75.815.6
OPP/EFO + kNN76.022.9
OPP/EFO + IF72.515.0
MOMENT51.91.7
TS2Vec50.11.9

CT + Stats + kNN leads on both metrics. On original cases, its AUROC reaches 87.1% versus 82.3% for Raw + kNN. On noise cases, it drops to 75.7% versus 79.7%. Shape-based detection is promising but noise-sensitive. The selected original cases below report AUROC / AUPRC in percent using kNN.

DatasetRawStatsCTCT+StatsOPPEFO
WalkingAceleration16.0 / 2.749.5 / 4.7100.0 / 98.999.5 / 93.5100.0 / 98.9100.0 / 98.9
InternalBleeding921.9 / 2.122.3 / 2.199.9 / 96.099.9 / 96.099.9 / 96.099.9 / 96.0
PowerDemand248.7 / 2.297.5 / 72.1100.0 / 99.5100.0 / 99.5100.0 / 99.5100.0 / 99.5
apneaecg353.5 / 0.467.1 / 0.6100.0 / 97.3100.0 / 96.9100.0 / 97.7100.0 / 97.7
TkeepForthMARS14.6 / 0.986.4 / 32.199.9 / 94.699.9 / 91.399.9 / 94.699.9 / 94.6
CHARISfive (9812)32.0 / 0.399.3 / 22.0100.0 / 94.3100.0 / 94.3100.0 / 94.3100.0 / 94.3
2sddb4077.6 / 1.595.6 / 8.0100.0 / 95.299.9 / 90.999.0 / 28.399.0 / 28.3
3sddb4081.8 / 1.989.3 / 3.1100.0 / 95.999.9 / 90.899.2 / 29.699.2 / 29.6
gait372.2 / 2.980.0 / 4.295.3 / 35.0100.0 / 100.095.3 / 35.095.3 / 35.0

These representative cases demonstrate that ordinal and hierarchical structural patterns can be highly informative for anomaly detection, even without magnitude statistics. CT features achieve performance comparable to OPP/EFO on several datasets while substantially outperforming them on 2sddb40 and 3sddb40. Adding magnitude statistics further improves some cases but does not consistently benefit every dataset. These findings suggest that CT-based representations are particularly useful for detecting structural anomalies.

W3. Learned representations

We included RDST and MiniROCKET for classification, zero-shot pretrained TimesFM 2.5/3.0 for forecasting, pretrained MOMENT reconstruction and dataset-trained TS2Vec for anomaly detection. Neural methods used GPUs, while classical features used CPUs.

AuthorsOpenReview ↗

Request for Discussion Engagement

Thank you for your time and constructive feedback. We greatly appreciate your positive assessment of our work.

As the discussion period approaches its end, we would appreciate it if you could review our responses and newly added experimental results. In particular, we have:

  • W1: Extended scalability experiments to sequences of up to 50,000 observations, directly comparing CT-Miner with CT-Hash and demonstrating increasing speedups for longer patterns.
  • W2: Added classification, forecasting, and anomaly detection evaluations, including analyses of the limitations of CT representations and the effects of augmenting them with magnitude statistics.
  • W3: Expanded comparisons with modern learning-based methods, including MiniROCKET, RDST, TimesFM, MOMENT, and TS2Vec, to clarify CT-Miner's strengths and limitations.

We hope these additional experiments adequately address your concerns. We would appreciate it if you could confirm whether your concerns have been resolved or let us know if any questions remain.

We would be happy to provide further clarification.

Comments on the paper

1 comment

We thank the reviewers for their constructive feedback. We have revised the manuscript to clarify the endmarker convention, algorithmic correctness, and computational advantages of CT-Miner. We expanded our experiments with an incremental counting baseline, early-stopping ablations, detailed memory profiling, and scalability evaluations on sequences of up to 50,000 observations. We also added downstream evaluations on classification, forecasting, and anomaly detection, including comparisons with modern learning-based baselines, and further discussed the strengths and limitations of CT-based representations. All revisions are highlighted in blue in the updated manuscript for ease of reference.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.