PaperlayerView exact revision on arXiv

Source-linked AI summary

ChorusTIC: Training-Free Multivariate Time Series Classification via Chorus In-Context Learning

Juntao Fang, Shifeng Xie, Ruichu Cai, Shengji Zheng, Zijian Li, Keli Zhang, Lujia Pan, Themis Palpanas, Zhifeng Hao

arXiv:2608.24033v1cs.LGcs.AIstat.ML

TL;DR

Time series classification often still requires fitting a task-specific classifier for each target dataset, while multivariate channels may be encoded independently. ChorusTIC performs classification directly from labeled context across heterogeneous channel configurations, achieving strong full-context and low-label performance on UEA-30 and UCR-128 without target-specific classifier fitting.

  • Problem

    Time series classification commonly uses target-specific classifier fitting, while multivariate inputs are often processed channel-wise despite evidence involving temporal intervals and cross-channel interactions.

  • Method

    ChorusTIC combines episode-consistent RSSC, a shared dual-axis encoder, context-derived calibration, and leakage-protected in-context classification, pretrained on labeled synthetic episodes.

  • Results

    ChorusTIC achieves strong full-context and low-label performance on the complete UEA-30 and UCR-128 archives without target-specific classifier fitting, improving consistently as labeled context grows.

  • Takeaways & Limitations

    Cross-channel modeling and support-conditioned inference provide complementary mechanisms for classification across heterogeneous channel configurations.

Abstract

from arXiv · show

Time series classification underpins applications in healthcare, sensing, and industrial monitoring. Although time series foundation models support forecasting and transferable representation learning, classification still typically requires fitting a task-specific classifier on each target dataset, while individual channels of multivariate inputs are often encoded independently. We introduce ChorusTIC, a classification-native foundation model for in-context classification across heterogeneous channel configurations without target-task parameter updates. ChorusTIC combines episode-consistent Random Subchannel Slot Concatenation with a shared dual-axis encoder to model temporal and cross-channel interactions and map variable channel configurations into a fixed-width representation independent of the original channel count. It then calibrates feature axes using context-derived distributions and predicts query labels through leakage-protected in-context learning. We pretrain ChorusTIC solely on synthetic labeled episodes comprising context and query sets that share a task background, with classes distinguished by sparse temporal or cross-channel rules. Evaluations on the complete UEA-30 and UCR-128 archives show strong full-context and low-label performance without target-specific classifier fitting.

Introduction

ChorusTIC addresses the need for classification models that capture temporal and cross-channel evidence while avoiding target-dataset classifier fitting. It combines heterogeneous-channel representations, context-conditioned calibration, and leakage-protected in-context prediction, pretrained on synthetic labeled episodes.

  • Multivariate classification may depend on localized temporal evidence, cross-variable interactions, sensor correlations, and temporal lead-lag relationships.
  • Representation-transfer pipelines typically fit a task-specific classifier for each target dataset after encoding samples.They remain sensitive to representation-layer choice, token aggregation, and downstream classifier selection.
  • ChorusTIC uses Random Subchannel Slot Concatenation and a shared dual-axis encoder to model temporal and within-group cross-channel interactions.Fixed-order slot concatenation produces a representation whose dimensionality is independent of the original channel count.
  • A leakage-protected in-context classifier predicts query labels from labeled context without target-task parameter updates.
  • Synthetic labeled episodes share a task-level temporal background while classes differ through sparse temporal or cross-channel rules.The contributions also include instance-level variation and episode-wise label permutation.

Related Work

Prior work has largely emphasized forecasting, transferable representations, or univariate in-context classification. ChorusTIC extends synthetic episodic pretraining and in-context classification toward multivariate tasks with cross-channel structure.

  • Time series foundation models support zero-shot prediction, limited-observation adaptation, and task-specific fine-tuning, whereas classification commonly uses representation transfer.
  • Directly inferring query labels for unseen classification tasks without target-specific optimization remains limited.
  • Existing in-context time series classifiers are primarily developed or evaluated for univariate classification.Other approaches use general-purpose language models with textual tables or channel-wise exemplars.
  • ChorusTIC uses a multivariate episodic prior in which context and query samples share a task background and sparse temporal or cross-channel rules determine class identity.

Method

ChorusTIC processes variable-channel episodes through signal-level slot-based encoding, task-level context calibration, and leakage-protected in-context classification. Synthetic episodes align pretraining with this support-conditioned deployment protocol.

  • Problem Formulation and Model Overview: The model computes signal representations, calibrates them with context, and produces query logits through a leakage-protected classifier.Its overall computation is H = Rψ(Xτ; I), U = Aϕ(H; Nc), and Oq = Gθ(U, Y c; Nc).
  • RSSC-Based Multivariate Representation: RSSC shares one episode-level channel-to-slot assignment across context and query samples before fixed-order slot composition.This maps variable channel sets to a representation width independent of the original channel count.
  • RSSC-Based Multivariate Representation: When channels are insufficient to fill slots, RSSC concatenates independently permuted channel lists and reuses observed channels instead of inserting zero-valued slots.
  • RSSC-Based Multivariate Representation: Within each sampled group, the dual-axis encoder models temporal dependencies within slots and interactions across aligned slots.Groups are later integrated by row-wise interaction rather than jointly processed by channel-axis attention.
  • Task-Level Calibration and ICL: Column Distribution Modeling builds context-only feature-axis references, while row-wise interaction processes each sample independently.This prevents query samples from modifying reference distributions or communicating with one another.
  • Task-Level Calibration and ICL: Labels enter only context tokens, and the ICL Transformer uses context tokens as keys and values while queries retain their own representations.Queries are predicted in parallel without direct query-to-query information flow.
  • Labeled Episodic Pretraining: Pretraining samples labeled episodes with shared task backgrounds, variable task settings, and classes defined by temporal or cross-channel rules.Univariate and multivariate tasks are sampled with probabilities 0.2 and 0.8, respectively.

Experiments

Experiments evaluate ChorusTIC across complete multivariate and univariate archives, frozen-feature and in-context baselines, and limited-label protocols. ChorusTIC achieves the strongest reported aggregate performance while avoiding target-specific parameter updates, with gains persisting as labeled context varies.

  • Multivariate classification on UEA: ChorusTIC achieves the highest average accuracy and best average rank on the complete UEA-30 archive without target-specific parameter updates.It improves average accuracy by approximately 2.51% over MantisV2+LR and reduces average rank from 4.02 to 3.57.
  • Multivariate classification on UEA: ChorusTIC improves average accuracy by approximately 6.34% over TabICLv2 and 10.62% over TabICL on UEA-30.These generic ICL baselines vectorize multivariate inputs without explicit temporal or aligned cross-channel modeling.
  • Univariate classification on UCR: ChorusTIC achieves the highest average accuracy and best average rank on the complete UCR-128 archive without target-specific classifier fitting.Relative to MantisV2+LR, it improves average accuracy by approximately 1.41% and reduces average rank from 5.50 to 4.43.
  • Univariate classification on UCR: On UCR-128, ChorusTIC improves average accuracy by approximately 1.44% over TIC-FM, 2.51% over TiCT, and 2.89% over TabICLv2.The same pretrained model therefore reports strong performance in both single-channel and multivariate settings without target-specific optimization.
  • Fixed-shot performance: ChorusTIC achieves the highest average accuracy with five and ten labeled examples per class, improving over the strongest baselines by approximately 1.11% and 4.06%, respectively.Against TabICL, the improvements are approximately 5.72% and 6.88% at five and ten shots.
  • Scaling with labeled context: ChorusTIC ranks first across 20%–60% labeled context on UEA-30, while average accuracy rises from 64.81% to 69.13%.Its relative improvement over the strongest competing method at each fraction ranges from 0.66% to 2.85%.

Conclusion

ChorusTIC performs support-conditioned univariate and multivariate time series classification without per-dataset classifier fitting, combining channel-aware encoding with leakage-protected in-context inference. Across UCR-128 and UEA-30, it achieves strong full-context and low-label performance, improving as labeled context grows, while motivating broader evaluation protocols.

  • Conclusion: ChorusTIC combines RSSC, a shared dual-axis encoder, context-derived calibration, and leakage-protected ICL to predict query labels without target-specific classifier fitting.The model is designed for heterogeneous channel configurations and models both temporal and cross-channel interactions.
  • Related Work: Existing TSFM classification pipelines typically transfer representations but still fit a task-specific classifier on every target dataset.ChorusTIC instead targets classification-native, parameter-update-free deployment.
  • Conclusion: Performance improves consistently as labeled context grows, indicating effective support-conditioned inference beyond the fixed-shot regime.The results address both fixed-shot and proportional low-label settings.
  • Conclusion: Cross-channel modeling and support-conditioned inference provide complementary mechanisms for classification across heterogeneous channel configurations.This conclusion connects the model’s channel-aware architecture with its context-based prediction strategy.
  • Related Work: TIC-FM, TiCT, RocketPFN, TableTime, and FETA offer related training-free or in-context routes, but differ in encoder design, modality, or multivariate handling.The cited methods include latent-memory, synthetic end-to-end, tabular-feature, textual-table, and channel-wise language-model approaches.

B Detailed Method

The detailed method constructs fixed-width, channel-aware representations through episode-consistent RSSC, patch tokenization, and shared dual-axis processing, then performs context-calibrated, leakage-protected in-context classification. Synthetic episodes align pretraining with deployment by sharing temporal backgrounds while varying sparse temporal or cross-channel class rules and instance nuisance transformations.

  • B.1 RSSC Group Construction: RSSC assigns channels to episode-consistent group-slot positions shared across context and query samples, reusing observed channels when necessary.Repeated slots remain valid observations rather than being replaced by zero-valued slots.
  • B.2 Patch Tokenization: Each selected channel is resampled to length L0 and divided into M non-overlapping patches of length w = L0/M.Patch features incorporate normalized values, first differences, and local statistics.
  • B.3 Shared Dual-Axis Encoder: The temporal block attends across patches within each slot, while the channel block attends across slots at aligned patch positions.A shared summary-token readout produces slot representations conditioned on other sampled channels in the group.
  • B.4 Fixed-Dimensional RSSC Composition: For fixed G, S, and ds, the representation dimension dh is independent of the original channel count C.RSSC preserves multiple sampled channel views without requiring globally aligned channel identities.
  • B.5 Context-Derived Calibration: Column Distribution Modeling processes feature axes independently using context-derived information, while a row Transformer models interactions among feature axes.The row Transformer is applied independently to each sample row, preventing cross-sample information flow at that stage.
  • B.6 Leakage-Protected In-Context Classification: The in-context classifier injects labels only into context tokens, making context the sole source of keys and values for query prediction.Query tokens do not serve as keys or values, preventing query-to-query communication and label leakage.
  • B.8 Complete Synthetic Episodic Prior: Episodes vary task type, rule family, difficulty, class proportions, and channel count, with multivariate episodes sampled over 2 ≤ C ≤ 10 channels.Every active class must be represented in the context set.
  • B.8 Complete Synthetic Episodic Prior: All classes in an episode share a temporal background, while sparse temporal or cross-channel rules determine class identity.Rule families include motifs, polarity, position, informative-channel selection, relative delays, phase relationships, and correlation regimes.

B.10 Hierarchical Extension for Many-Class Tasks

For tasks exceeding the decoder’s native class capacity, ChorusTIC decomposes classification into hierarchical in-context decisions. A balanced tree groups classes into bounded-child nodes, and the original class probability is formed along the selected path without target-task parameter updates.

  • B.10 Hierarchical Extension for Many-Class Tasks: When K > Kmax, a balanced tree organizes original classes into internal nodes with at most Kmax children.The tree’s leaves correspond to the original classes.
  • B.10 Hierarchical Extension for Many-Class Tasks: At each internal node, context labels are replaced by local child-group indices for the corresponding descendant classes.The model predicts the next child group using the context samples associated with that node.
  • B.10 Hierarchical Extension for Many-Class Tasks: A class probability is obtained by combining the in-context decisions along the class’s path from the root to its leaf.Each path node selects one child group toward the target class.
  • B.10 Hierarchical Extension for Many-Class Tasks: The hierarchical procedure converts many-class prediction into a sequence of native-capacity in-context decisions without target-task parameter updates.This extends the decoder to class counts beyond Kmax while preserving training-free inference.

C.1 Datasets and Evaluation Splits

The study evaluates ChorusTIC on official UEA-30 and UCR-128 train/test splits, using labeled training data as context and test data as queries without updating target-task parameters. Evaluation includes full-context and fixed-shot settings, standardized preprocessing, frozen baselines, macro-averaged metrics, and paired statistical testing.

  • UEA-30 and UCR-128 are evaluated using their official train/test splits, with no datasets or labels modified.
  • Full-context evaluation provides the complete training split as labeled context and the complete test split as the query set, without updating ChorusTIC parameters.
  • Fixed-shot evaluation samples 5 or 10 labeled examples per class, averages five independently sampled support sets, and excludes datasets only when a class lacks the requested shot count.
  • UCR series are represented with a singleton channel axis, while UEA channels retain their multivariate organization without flattening, concatenation, selection, or truncation.
  • Inputs are resampled to L0 = 512, missing UEA values are zero-filled before interpolation, and preprocessing is applied consistently to train and test splits without test labels or statistics.
  • Archive-level results use unweighted average accuracy and average rank, while paired Wilcoxon tests compare per-dataset accuracies with Holm correction within each archive.

D.1 Evaluation Details

Evaluation compares ChorusTIC with time-series ICL methods, generic ICL controls, and frozen foundation-model baselines under full-context and low-label protocols. Metrics summarize archive-level accuracy, rank, and paired win/tie/loss outcomes.

  • UEA-30 and UCR-128 use training splits as labeled context and test splits as queries, with no target-task parameter updates.
  • The comparison includes time-series ICL methods, generic vectorized ICL controls, and frozen TSFM backbones using their released implementations and target-classifier protocols.
  • Reported metrics are unweighted average accuracy, average rank, and ChorusTIC-perspective win/tie/loss counts; low-label results average five sampled context sets.

D.2 Archive-Level Results

ChorusTIC achieves strong archive-level performance across multivariate UEA-30 and univariate UCR-128 classification while retaining training-free deployment. Its UEA results also highlight the value of native multivariate encoding relative to generic ICL controls.

  • ChorusTIC is the only evaluated method combining training-free deployment with native multivariate encoding on UEA-30.
  • UEA-30 Table D.1 reports target-fit status, archive-level accuracy and rank, and win/tie/loss counts from ChorusTIC’s perspective.
  • ChorusTIC achieves the best aggregate accuracy and rank on UCR-128 without fitting a target-specific classifier.

D.3 Per-Dataset Results

Per-dataset tables expose heterogeneity behind the archive averages while preserving the same training-free protocol. Statistical tests report corrected significance for several baseline comparisons in both archives.

  • UEA-30 per-dataset accuracies vary across heterogeneous multivariate tasks while confirming ChorusTIC’s best archive-level average accuracy and rank.
  • UCR-128 per-dataset results show variation across all 128 tasks while retaining ChorusTIC’s training-free classification protocol.
  • On UCR-128, corrected differences remain significant for six baselines, including NuTime+RF with W/T/L = 121/1/6 and adjusted p = 4.90 × 10^-30.
Loading 2608.24033v1…