📈 Time Series¶

🤖 AAAI2026 · 35 paper notes

A Theoretical Analysis of Detecting Large Model-Generated Time Series: This work presents the first theoretical framework for detecting time series large model (TSLM)-generated content. By establishing the Contraction Hypothesis, it reveals that TSLM-generated sequences exhibit exponentially decaying uncertainty under recursive forecasting. Based on this insight, the proposed UCE detector achieves an in-distribution AUROC of 0.855 across 32 datasets, substantially outperforming 10 text-detection baselines.
A Unified Shape-Aware Foundation Model for Time Series Classification: This paper proposes UniShape — a foundation model for time series classification that adaptively aggregates multi-scale discriminative subsequences (shapelets) via a shape-aware adapter, and learns transferable shapelet representations at both instance and shape levels through prototype-based contrastive pretraining. With only 3.1M parameters, UniShape achieves state-of-the-art performance on 128 UCR datasets (average accuracy 87.08%) while providing strong classification interpretability.
AirDDE: Multifactor Neural Delay Differential Equations for Air Quality Forecasting: The first framework to introduce Neural Delay Differential Equations (NDDE) into air quality forecasting. By incorporating a memory-augmented attention module and a physics-guided delay evolution function, it models delay effects in the continuous-time propagation of pollutants, achieving an average MAE reduction of 8.79% across three datasets.
iTimER: Reconstruction Error-Guided Irregularly Sampled Time Series Representation Learning: This paper proposes iTimER, which leverages the model's own reconstruction error distribution as a learning signal. By estimating the error distribution from observed points and sampling from it to generate pseudo-observations at unobserved timestamps, the method aligns the error distributions of observed and pseudo-observed regions via Wasserstein distance combined with contrastive learning, achieving state-of-the-art performance on classification, interpolation, and forecasting tasks for irregularly sampled time series.
C3RL: Rethinking the Combination of Channel-independence and Channel-mixing from Representation Learning: This paper proposes C3RL, a SimSiam-based contrastive learning framework that treats channel-independence (CI) and channel-mixing (CM) strategies as two transposed views of the same data to construct positive pairs. By jointly optimizing representation learning and forecasting through a Siamese network, C3RL improves the best-performance rate of CI models from 43.6% to 81.4% and CM models from 23.8% to 76.3%.
Coherent Multi-Agent Trajectory Forecasting in Team Sports with CausalTraj: This paper proposes CausalTraj — a temporally causal, likelihood-based multi-agent trajectory forecasting model that autoregressively models spatio-temporal interactions among agents step by step. CausalTraj achieves state-of-the-art results on joint metrics (minJADE/minJFDE) across NBA, basketball, and football datasets while maintaining competitive per-agent accuracy.
CometNet: Contextual Motif-guided Long-term Time Series Forecasting: This paper proposes CometNet, which extracts recurrently occurring "contextual motifs" from the full historical sequence to construct a motif library, and employs a motif-guided MoE architecture to dynamically associate the current window with relevant motifs for prediction. This approach breaks the receptive field bottleneck imposed by limited look-back windows and achieves significant improvements over state-of-the-art methods such as TimeMixer++ and iTransformer on 8 datasets.
Counterfactual Explainable AI (XAI) Method for Deep Learning-Based Multivariate Time Series Classification: This paper proposes CONFETTI, a multi-objective counterfactual explanation method for multivariate time series (MTS) classification. By combining Class Activation Map (CAM)-guided subsequence extraction with NSGA-III multi-objective optimization, CONFETTI achieves an optimal balance among prediction confidence, sparsity, and proximity, outperforming existing methods across 7 UEA benchmark datasets.
DeepBooTS: Dual-Stream Residual Boosting for Drift-Resilient Time-Series Forecasting: This paper proposes DeepBooTS, which leverages bias-variance decomposition theory to demonstrate that weighted ensembling reduces variance and thereby mitigates concept drift. The method introduces a dual-stream residual-decreasing boosting architecture in which each block corrects the residual of the preceding block, achieving an average improvement of 15.8% across multiple datasets.
Detecting the Future: All-at-Once Event Sequence Forecasting with Horizon Matching: This paper proposes DEF (Detection-based Event Forecasting), which draws on the set-matching idea from DETR in object detection and employs the Hungarian algorithm to align predicted and ground-truth event sequences, achieving high-accuracy and high-diversity long-horizon event forecasting with state-of-the-art results on five datasets.
Finding Time Series Anomalies using Granular-ball Vector Data Description: This paper proposes the Granular-ball One-Class Network (GBOC), which adaptively constructs density-guided Granular-ball Vector Data Descriptions (GVDD) in the latent space. By replacing traditional clustering or single-hypersphere assumptions, GBOC enables flexible modeling of normal time series behavior and robust anomaly detection.
FreqCycle: A Multi-Scale Time-Frequency Analysis Method for Time Series Forecasting: This paper proposes the FreqCycle framework, which explicitly learns shared periodic patterns via the FECF module, enhances mid-to-high frequency energy contributions via the SFPL module, and extends to MFreqCycle for handling coupled multi-periodicity. The framework achieves an optimal balance of SOTA performance and efficiency across 7 benchmarks.
GAICo: A Deployed and Extensible Framework for Evaluating Diverse and Multimodal Generative AI Outputs: This paper presents GAICo (Generative AI Comparator), a deployed, extensible, open-source Python library that provides a unified reference-based evaluation framework for text, structured data (planning sequences, time series), and multimedia (images, audio), supporting multi-model comparison, visualization, and report generation.
Harmonic Dataset Distillation for Time Series Forecasting: This paper proposes HDT (Harmonic Dataset Distillation for Time Series Forecasting), which decomposes time series into sinusoidal bases via FFT and aligns the core periodic structure of synthetic and real data through Harmonic Matching in the frequency domain, achieving strong cross-architecture generalization and favorable scalability for time series dataset distillation.
HN-MVTS: HyperNetwork-based Multivariate Time Series Forecasting: This paper proposes HN-MVTS, which employs a HyperNetwork to generate channel-specific weights for the final prediction layer, striking a balance between channel-independent (CI) and channel-dependent (CD) modeling. As a plug-and-play module, it improves forecasting accuracy of various backbone models including DLinear, PatchTST, and TSMixer without incurring additional inference overhead.
HydroDCM: Hydrological Domain-Conditioned Modulation for Cross-Reservoir Inflow Prediction: This paper proposes HydroDCM, the first framework to introduce Domain Generalization (DG) into hydrological forecasting. It constructs pseudo-domain labels from spatial meta-attributes to guide adversarial learning for invariant feature extraction, then employs a FiLM adapter to modulate features conditioned on the target reservoir's geographical information, enabling cross-domain inflow prediction for unseen reservoirs.
IdealTSF: Can Non-Ideal Data Contribute to Enhancing Time Series Forecasting?: IdealTSF is a three-stage progressive framework that (1) uses negative sample pre-training on synthetic non-ideal data to enhance robustness, (2) trains on repaired positive samples to learn underlying trends, and (3) applies the ECOS optimizer to guide parameters toward flat minima — achieving approximately 10% MSE improvement on time series data containing noise and missing values.
Interpreting Fedspeak with Confidence: A LLM-Based Uncertainty-Aware Framework Guided by Monetary Policy Transmission Paths: This paper proposes an LLM-based uncertainty-aware framework for interpreting Fedspeak (Federal Reserve language). The framework enhances inputs through domain reasoning along monetary policy transmission paths, and introduces a dynamic uncertainty decoding module to quantify prediction confidence (Perceptual Uncertainty = Environmental Ambiguity × Cognitive Risk), achieving SOTA performance on FOMC monetary policy stance analysis.
LoReTTA: A Low Resource Framework To Poison Continuous Time Dynamic Graphs: This paper proposes LoReTTA, a two-stage adversarial poisoning attack framework that requires no surrogate model. It first sparsifies high-influence edges via 16 temporal importance metrics, then replaces them with adversarial edges using a degree-preserving negative sampling algorithm. Across 4 datasets × 4 TGNN models, LoReTTA achieves an average performance degradation of 29.47%, while evading 4 anomaly detection systems and resisting 4 defense methods.
M2FMoE: Multi-Resolution Multi-View Frequency Mixture-of-Experts for Extreme-Adaptive Time Series Forecasting: This paper proposes M2FMoE, a framework that models both regular and extreme temporal patterns via frequency-domain Mixture-of-Experts from dual Fourier and wavelet perspectives. It incorporates a cross-view shared frequency-band splitter to align semantic correspondence across domains, multi-resolution adaptive fusion to capture multi-scale information, and temporal gated integration to combine short- and long-term features. On five hydrological extreme event datasets, M2FMoE surpasses all state-of-the-art methods — including label-supervised approaches — without requiring any extreme event labels, achieving an average RMSE improvement of 22.3%.
Mask the Redundancy: Evolving Masking Representation Learning for Multivariate Time-Series Clustering: This paper proposes EMTC, a framework that dynamically masks redundant timestamps via Importance-aware Variate-wise Masking (IVM), combined with Multi-Endogenous Views (MEV) generation and cluster-guided contrastive learning, achieving an average F1 improvement of 4.85% across 15 MTS clustering benchmarks.
Mitigating Error Accumulation in Co-Speech Motion Generation via Global Rotation Diffusion and Multi-Level Constraints: This paper proposes GlobalDiff, a framework that, for the first time, performs diffusion-based generation in the global joint rotation space, fundamentally eliminating error accumulation in hierarchical forward kinematics. A three-level joint–bone–motion constraint scheme compensates for the structural priors lost under global representation. GlobalDiff achieves state-of-the-art performance on multi-speaker co-speech motion generation benchmarks, improving FGD by 46% over the previous best method.
Optimal Look-back Horizon for Time Series Forecasting in Federated Learning: This paper proposes a theoretical framework for selecting the optimal look-back horizon in federated time series forecasting. By introducing a Synthetic Data Generator (SDG) and an intrinsic space representation, the forecasting loss is decomposed into an irreducible Bayesian error and an approximation error. The paper proves that the total loss is unimodal with respect to the horizon length, and establishes that the minimum sufficient window is the optimal solution.
Predicting the Future by Retrieving the Past: This paper proposes PFRP (Predicting the Future by Retrieving the Past), which constructs a Global Memory Bank (GMB) to store historical patterns, trains an encoder via Predictive Contrastive Learning (PCL) for efficient retrieval, and dynamically integrates retrieved global predictions with any local forecasting model. PFRP achieves an average improvement of 8.4% in forecasting performance across 7 datasets.
ProbFM: Probabilistic Time Series Foundation Model with Uncertainty Decomposition: This work is the first to introduce Deep Evidential Regression (DER) with a Normal-Inverse-Gamma prior into a time series foundation model architecture, enabling epistemic-aleatoric uncertainty decomposition in a single forward pass. The practical value of uncertainty-aware trading strategies is validated on cryptocurrency forecasting.
ReCast: Reliability-aware Codebook Assisted Lightweight Time Series Forecasting: This paper proposes ReCast, which encodes time series into discrete embeddings via patch-level vector quantization. It introduces a dual-path architecture consisting of a quantization path (modeling regular structures) and a residual path (capturing irregular fluctuations), along with a reliability-aware codebook update strategy based on distributionally robust optimization (DRO). ReCast achieves state-of-the-art accuracy with a lightweight architecture across 8 datasets.
Revitalizing Canonical Pre-Alignment for Irregular Multivariate Time Series Forecasting: This paper is the first to argue that Canonical Pre-Alignment (CPA) should not be abandoned for Irregular Multivariate Time Series (IMTS) forecasting. It proposes KAFNet, which addresses the efficiency bottleneck of CPA via three modules—Pre-Convolution smoothing, Temporal Kernel Aggregation (TKA), and Frequency-domain Linear Attention (FLA)—achieving state-of-the-art accuracy on 4 IMTS benchmarks while reducing parameters by 7.2× and accelerating training/inference by 8.4×.
Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios: This paper proposes SpecFormer, a non-autoregressive draft model architecture that integrates unidirectional and bidirectional attention. By reducing reliance on large prefix trees and minimizing position-dependent parameters, SpecFormer achieves consistent LLM inference acceleration in large-batch scenarios.
SELDON: Supernova Explosions Learned by Deep ODE Networks: This paper proposes SELDON, a continuous-time VAE combining a masked GRU-ODE encoder, an implicit Neural ODE propagator, and an interpretable Gaussian basis function decoder, designed for sparse and irregularly sampled astronomical light curve prediction. SELDON outperforms baseline methods in accurate multi-band flux prediction using only 20% of observed data.
Sonnet: Spectral Operator Neural Network for Multivariable Time Series Forecasting: This paper proposes Sonnet, which maps inputs to the time-frequency domain via learnable wavelet transforms, introduces multivariate coherence-based attention (MVCA) to model inter-variable dependencies, and employs a Koopman operator for stable temporal evolution forecasting. Sonnet achieves state-of-the-art performance on 34 out of 47 forecasting tasks, reducing average MAE by 2.2%.
Task-Aware Retrieval Augmentation for Dynamic Recommendation: This paper proposes TarDGR, a framework that automatically constructs training data via a task-aware evaluation mechanism, trains a Graph Transformer to assess the task relevance of historical subgraphs, and retrieves and integrates task-relevant subgraphs at inference time to enhance temporal generalization in recommendation.
Towards Non-Stationary Time Series Forecasting with Temporal Stabilization and Frequency Differencing: This paper proposes DTAF, a dual-branch framework that extracts and removes heterogeneous non-stationary patterns via a non-stationary MoE filter in the temporal domain, tracks frequency drift via spectral differencing in the frequency domain, and fuses complementary information from both domains through dual-branch attention for robust non-stationary time series forecasting.
Transparent Networks for Multivariate Time Series: This paper proposes GATSM (Generalized Additive Time Series Model), a transparent neural network for time series that employs weight-sharing feature networks to learn feature representations and masked multi-head attention to capture temporal patterns. GATSM achieves performance comparable to black-box models such as Transformers while maintaining full interpretability.
Urban Incident Prediction with Graph Neural Networks: Integrating Government Ratings and Crowdsourced Reports: This paper proposes URBAN, a multi-view multi-output GNN model that jointly leverages sparse but unbiased government inspection rating data and dense but biased crowdsourced report data to predict the true latent state of urban incidents. Validated on 9.6M+ reports and 1M+ inspections in New York City, the model achieves a 5.3× higher prediction correlation than using report data alone.
XLinear: A Lightweight and Accurate MLP-Based Model for Long-Term Time Series Forecasting with Exogenous Inputs: This paper proposes XLinear, a lightweight time series forecasting model based on MLP with sigmoid gating. Through a global token mechanism, it efficiently integrates endogenous and exogenous variable information, achieving an optimal accuracy–efficiency trade-off across 12 datasets.