Pre-trained Large Language Models (LLMs) encapsulate large amounts of knowledge and take enormous amounts of compute to train. We make use of this resource, together with the observation that LLMs are able to transfer knowledge and performance from one domain or even modality to another seemingly-unrelated area, to help with multivariate demand time series forecasting. Attention in transformer-based methods requires something worth attending to -- more than just samples of a time-series. We explore different methods to map multivariate input time series into the LLM token embedding space. In particular, our novel multivariate patching strategy to embed time series features into decoder-only pre-trained Transformers produces results competitive with state-of-the-art time series forecasting models. We also use recently-developed weight-based diagnostics to validate our findings.
Current deep regression models usually learn in point-wise way that treat each sample as an independent input, neglecting the relative ordering among different data. Consequently, the regression model could neglect the data 's interrelationships, potentially resulting in suboptimal performance. Moreover, the existence of aleatoric uncertainty in the training data may drive the model to capture non-generalizable patterns, contributing to increased overfitting. To address these issues, we propose a novel adaptive pairwise learning framework (AdaPRL) for regression tasks which leverages the relative differences between data points and integrates with deep probabilistic models to quantify the uncertainty associated with the predictions. Additionally, we adapt AdaPRL for applications in multi-task learning and multivariate time series forecasting. Extensive experiments with several real-world regression datasets including recommendation systems, age estimation, time series forecasting, natural language understanding, finance, and industry datasets show that AdaPRL is compatible with different backbone networks in various tasks and achieves state-of-the-art performance on the vast majority of tasks, highlighting its notable potential including enhancing prediction accuracy and ranking ability, increasing generalization capability, improving robustness to noisy data, improving resilience to reduced data, and enhancing interpretability, etc.
We introduce an interpretable deep learning model for multivariate time series forecasting that prioritizes both predictive performance and interpretability - key requirements for understanding complex physical phenomena. Our model not only matches but often surpasses existing interpretability methods, achieving this without compromising accuracy. Through extensive experiments, we demonstrate its ability to identify the most relevant time series and lags that contribute to forecasting future values, providing intuitive and transparent explanations for its predictions. To minimize the need for manual supervision, the model is designed so one can robustly determine the optimal window size that captures all necessary interactions within the smallest possible time frame. Additionally, it effectively identifies the optimal model order, balancing complexity when incorporating higher-order terms. These advancements hold significant implications for modeling and understanding dynamic systems, making the model a valuable tool for applied and computational physicists.
Among the existing Transformer-based multivariate time series forecasting methods, iTransformer, which treats each variable sequence as a token and only explicitly extracts cross-variable dependencies, and PatchTST, which adopts a channel-independent strategy and only explicitly extracts cross-time dependencies, both significantly outperform most Channel-Dependent Transformer that simultaneously extract cross-time and cross-variable dependencies. This indicates that existing Transformer-based multivariate time series forecasting methods still struggle to effectively fuse these two types of information. We attribute this issue to the dynamic time lags in the causal relationships between different variables. Therefore, we propose a new multivariate time series forecasting Transformer, Sensorformer, which first compresses the global patch information and then simultaneously extracts cross-variable and cross-time dependencies from the compressed representations. Sensorformer can effectively capture the correct inter-variable correlations and causal relationships, even in the presence of dynamic causal lags between variables, while also reducing the computational complexity of pure cross-patch self-attention from $O(D^2 \cdot Patch\_num^2 \cdot d\_model)$ to $O(D^2 \cdot Patch\_num \cdot d\_model)$. Extensive comparative and ablation experiments on 9 mainstream real-world multivariate time series forecasting datasets demonstrate the superiority of Sensorformer. The implementation of Sensorformer, following the style of the Time-series-library and scripts for reproducing the main results, is publicly available at https://github.com/BigYellowTiger/Sensorformer
In long-term time series forecasting, Transformer-based models have achieved great success, due to its ability to capture long-range dependencies. However, existing models face challenges in identifying critical components for prediction, leading to limited interpretability and suboptimal performance. To address these issues, we propose the Inverted Seasonal-Trend Decomposition Transformer (Ister), a novel Transformer-based model for multivariate time series forecasting. Ister decomposes time series into seasonal and trend components, further modeling multi-periodicity and inter-series dependencies using a Dual Transformer architecture. We introduce a novel Dot-attention mechanism that improves interpretability, computational efficiency, and predictive accuracy. Comprehensive experiments on benchmark datasets demonstrate that Ister outperforms existing state-of-the-art models, achieving up to 10% improvement in MSE. Moreover, Ister enables intuitive visualization of component contributions, shedding lights on model's decision process and enhancing transparency in prediction results.
In recent years, the application of transformer-based models in time-series forecasting has received significant attention. While often demonstrating promising results, the transformer architecture encounters challenges in fully exploiting the temporal relations within time series data due to its attention mechanism. In this work, we design eXponential Patch (xPatch for short), a novel dual-stream architecture that utilizes exponential decomposition. Inspired by the classical exponential smoothing approaches, xPatch introduces the innovative seasonal-trend exponential decomposition module. Additionally, we propose a dual-flow architecture that consists of an MLP-based linear stream and a CNN-based non-linear stream. This model investigates the benefits of employing patching and channel-independence techniques within a non-transformer model. Finally, we develop a robust arctangent loss function and a sigmoid learning rate adjustment scheme, which prevent overfitting and boost forecasting performance. The code is available at the following repository: https://github.com/stitsyuk/xPatch.
Irregularly sampled multivariate time series (ISMTS) are prevalent in reality. Due to their non-uniform intervals between successive observations and varying sampling rates among series, the channel-independent (CI) strategy, which has been demonstrated more desirable for complete multivariate time series forecasting in recent studies, has failed. This failure can be further attributed to the sampling sparsity, which provides insufficient information for effective CI learning, thereby reducing its capacity. When we resort to the channel-dependent (CD) strategy, even higher capacity cannot mitigate the potential loss of diversity in learning similar embedding patterns across different channels. We find that existing work considers CI and CD strategies to be mutually exclusive, primarily because they apply these strategies to the global channel. However, we hold the view that channel strategies do not necessarily have to be used globally. Instead, by appropriately applying them locally and globally, we can create an opportunity to take full advantage of both strategies. This leads us to introduce the Channel Harmony ISMTS Transformer (TimeCHEAT), which utilizes the CD locally and the CI globally. Specifically, we segment the ISMTS into sub-series level patches. Locally, the CD strategy aggregates information within each patch for time embedding learning, maximizing the use of relevant observations while reducing long-range irrelevant interference. Here, we enhance generality by transforming embedding learning into an edge weight prediction task using bipartite graphs, eliminating the need for special prior knowledge. Globally, the CI strategy is applied across patches, allowing the Transformer to learn individualized attention patterns for each channel. Experimental results indicate our proposed TimeCHEAT demonstrates competitive SOTA performance across three mainstream tasks.
Time series forecasting is a crucial challenge with significant applications in areas such as weather prediction, stock market analysis, and scientific simulations. This paper introduces an embedded decomposed transformer, 'EDformer', for multivariate time series forecasting tasks. Without altering the fundamental elements, we reuse the Transformer architecture and consider the capable functions of its constituent parts in this work. Edformer first decomposes the input multivariate signal into seasonal and trend components. Next, the prominent multivariate seasonal component is reconstructed across the reverse dimensions, followed by applying the attention mechanism and feed-forward network in the encoder stage. In particular, the feed-forward network is used for each variable frame to learn nonlinear representations, while the attention mechanism uses the time points of individual seasonal series embedded within variate frames to capture multivariate correlations. Therefore, the trend signal is added with projection and performs the final forecasting. The EDformer model obtains state-of-the-art predicting results in terms of accuracy and efficiency on complex real-world time series datasets. This paper also addresses model explainability techniques to provide insights into how the model makes its predictions and why specific features or time steps are important, enhancing the interpretability and trustworthiness of the forecasting results.
Multivariate time series forecasting is crucial for various applications, such as financial investment, energy management, weather forecasting, and traffic optimization. However, accurate forecasting is challenging due to two main factors. First, real-world time series often show heterogeneous temporal patterns caused by distribution shifts over time. Second, correlations among channels are complex and intertwined, making it hard to model the interactions among channels precisely and flexibly. In this study, we address these challenges by proposing a general framework called DUET, which introduces dual clustering on the temporal and channel dimensions to enhance multivariate time series forecasting. First, we design a Temporal Clustering Module (TCM) that clusters time series into fine-grained distributions to handle heterogeneous temporal patterns. For different distribution clusters, we design various pattern extractors to capture their intrinsic temporal patterns, thus modeling the heterogeneity. Second, we introduce a novel Channel-Soft-Clustering strategy and design a Channel Clustering Module (CCM), which captures the relationships among channels in the frequency domain through metric learning and applies sparsification to mitigate the adverse effects of noisy channels. Finally, DUET combines TCM and CCM to incorporate both the temporal and channel dimensions. Extensive experiments on 25 real-world datasets from 10 application domains, demonstrate the state-of-the-art performance of DUET.
The progress of deep-learning-based forecasting architectures is evident through their expanding parameter configurations. However, the need for rapid online decision making in practical scenarios calls for an alternative strategy, highlighting the necessity for networks that are not only adaptive but also efficient in real-time operations. This shift is critical as we confront three principal challenges in deep-learning-based forecasting frameworks: (i) the inherent limitations of transformers, which, despite their attempts to preserve ordering information, the temporal information loss due to the permutation-invariant nature of self-attention mechanisms is inevitable, (ii) the inefficacy of linear models in capturing the dynamic interactions within swiftly evolving signals; and (iii) the incapacity of tree-based approaches to extrapolating beyond values present in the training set. In response to these challenges, we introduce LTBoost, an innovative boosted hybrid of linear and tree-based ensemble gradient algorithms tailored for long-term time series forecasting (LTSF) tasks, scalable to high data dimensions. LTBoost employs a dual strategy, beginning with a linear regression model to capture trends and extrapolate beyond known data, complemented by a robust nonlinear tree-based model that focuses on the residuals. This boosted hybrid approach not only addresses the challenges posed by existing models but also significantly improves forecast accuracy. The effectiveness of LTBoost is validated through empirical experiments conducted on nine well-established benchmark datasets, demonstrating superior performance and achieving state-of-the-art results in 32 out of 36 cases, measured by mean absolute error (MAE). Our findings also explore the impact of lag features and signal normalization techniques, demonstrating further improvements in predictive accuracy. This hybrid and highly effective approach highlights LTBoost's innovation and its resolution of specific forecasting challenges, setting the stage for its contribution to the field of time series forecasting, paving the way for its application in diverse real-world scenarios.
Transformers are now ubiquitous for sequence modeling tasks, but their extension to multi-dimensional data remains a challenge due to the quadratic cost of the attention mechanism. In this paper, we propose Higher-Order Transformers (HOT), a novel architecture designed to efficiently process data with more than two axes, i.e. higher-order tensors. To address the computational challenges associated with high-order tensor attention, we introduce a novel Kronecker factorized attention mechanism that reduces the attention cost to quadratic in each axis' dimension, rather than quadratic in the total size of the input tensor. To further enhance efficiency, HOT leverages kernelized attention, reducing the complexity to linear. This strategy maintains the model's expressiveness while enabling scalable attention computation. We validate the effectiveness of HOT on two high-dimensional tasks, including multivariate time series forecasting, and 3D medical image classification. Experimental results demonstrate that HOT achieves competitive performance while significantly improving computational efficiency, showcasing its potential for tackling a wide range of complex, multi-dimensional data.
Multivariate time series data provide a robust framework for future predictions by leveraging information across multiple dimensions, ensuring broad applicability in practical scenarios. However, their high dimensionality and mixing patterns pose significant challenges in establishing an interpretable and explicit mapping between historical and future series, as well as extracting long-range feature dependencies. To address these challenges, we propose a channel-time dual unmixing network for multivariate time series forecasting (named MTS-UNMixer), which decomposes the entire series into critical bases and coefficients across both the time and channel dimensions. This approach establishes a robust sharing mechanism between historical and future series, enabling accurate representation and enhancing physical interpretability. Specifically, MTS-UNMixers represent sequences over time as a mixture of multiple trends and cycles, with the time-correlated representation coefficients shared across both historical and future time periods. In contrast, sequence over channels can be decomposed into multiple tick-wise bases, which characterize the channel correlations and are shared across the whole series. To estimate the shared time-dependent coefficients, a vanilla Mamba network is employed, leveraging its alignment with directional causality. Conversely, a bidirectional Mamba network is utilized to model the shared channel-correlated bases, accommodating noncausal relationships. Experimental results show that MTS-UNMixers significantly outperform existing methods on multiple benchmark datasets. The code is available at https://github.com/ZHU-0108/MTS-UNMixers.
Industry 5.0 introduces new challenges for Long-term Time Series Forecasting (LTSF), characterized by high-dimensional, high-resolution data and high-stakes application scenarios. Against this backdrop, developing efficient and interpretable models for LTSF becomes a key challenge. Existing deep learning and linear models often suffer from excessive parameter complexity and lack intuitive interpretability. To address these issues, we propose DiPE-Linear, a Disentangled interpretable Parameter-Efficient Linear network. DiPE-Linear incorporates three temporal components: Static Frequential Attention (SFA), Static Temporal Attention (STA), and Independent Frequential Mapping (IFM). These components alternate between learning in the frequency and time domains to achieve disentangled interpretability. The decomposed model structure reduces parameter complexity from quadratic in fully connected networks (FCs) to linear and computational complexity from quadratic to log-linear. Additionally, a Low-Rank Weight Sharing policy enhances the model's ability to handle multivariate series. Despite operating within a subspace of FCs with limited expressive capacity, DiPE-Linear demonstrates comparable or superior performance to both FCs and nonlinear models across multiple open-source and real-world LTSF datasets, validating the effectiveness of its sophisticatedly designed structure. The combination of efficiency, accuracy, and interpretability makes DiPE-Linear a strong candidate for advancing LTSF in both research and real-world applications. The source code is available at https://github.com/wintertee/DiPE-Linear.
Variable Subset Forecasting (VSF) refers to a unique scenario in multivariate time series forecasting, where available variables in the inference phase are only a subset of the variables in the training phase. VSF presents significant challenges as the entire time series may be missing, and neither inter- nor intra-variable correlations persist. Such conditions impede the effectiveness of traditional imputation methods, primarily focusing on filling in individual missing data points. Inspired by the principle of feature engineering that not all variables contribute positively to forecasting, we propose Task-Oriented Imputation for VSF (TOI-VSF), a novel framework shifts the focus from accurate data recovery to directly support the downstream forecasting task. TOI-VSF incorporates a self-supervised imputation module, agnostic to the forecasting model, designed to fill in missing variables while preserving the vital characteristics and temporal patterns of time series data. Additionally, we implement a joint learning strategy for imputation and forecasting, ensuring that the imputation process is directly aligned with and beneficial to the forecasting objective. Extensive experiments across four datasets demonstrate the superiority of TOI-VSF, outperforming baseline methods by $15\%$ on average.
Understanding the relationship between cognition and intrinsic brain activity through purely data-driven approaches remains a significant challenge in neuroscience. Resting-state functional magnetic resonance imaging (rs-fMRI) offers a non-invasive method to monitor regional neural activity, providing a rich and complex spatiotemporal data structure. Deep learning has shown promise in capturing these intricate representations. However, the limited availability of large datasets, especially for disease-specific groups such as Alzheimer's Disease (AD), constrains the generalizability of deep learning models. In this study, we focus on multivariate time series forecasting of independent component networks derived from rs-fMRI as a form of data augmentation, using both a conventional LSTM-based model and the novel Transformer-based BrainLM model. We assess their utility in AD classification, demonstrating how generative forecasting enhances classification performance. Post-hoc interpretation of BrainLM reveals class-specific brain network sensitivities associated with AD.
Attention-based architectures have become ubiquitous in time series forecasting tasks, including spatio-temporal (STF) and long-term time series forecasting (LTSF). Yet, our understanding of the reasons for their effectiveness remains limited. In this work, we propose a novel pruning strategy, $\textbf{R}$eplace $\textbf{A}$ttention with $\textbf{M}$LP (RAM), that approximates the attention mechanism using only feedforward layers, residual connections, and layer normalization for temporal and/or spatial modeling in multivariate time series forecasting. Specifically, the Q, K, and V projections, the attention score calculation, the dot-product between the attention score and the V, and the final projection can be removed from the attention-based networks without significantly degrading the performance, so that the given network remains the top-tier compared to other SOTA methods. RAM achieves a $62.579\%$ reduction in FLOPs for spatio-temporal models with less than $2.5\%$ performance drop, and a $42.233\%$ FLOPs reduction for LTSF models with less than $2\%$ performance drop.
Transformer-based architectures have achieved remarkable success in natural language processing and computer vision. However, their performance in multivariate long-term forecasting often falls short compared to simpler linear baselines. Previous research has identified the traditional attention mechanism as a key factor limiting their effectiveness in this domain. To bridge this gap, we introduce LATST, a novel approach designed to mitigate entropy collapse and training instability common challenges in Transformer-based time series forecasting. We rigorously evaluate LATST across multiple real-world multivariate time series datasets, demonstrating its ability to outperform existing state-of-the-art Transformer models. Notably, LATST manages to achieve competitive performance with fewer parameters than some linear models on certain datasets, highlighting its efficiency and effectiveness.
In recent years, Transformer-based models (Transformers) have achieved significant success in multivariate time series forecasting (MTSF). However, previous works focus on extracting features either from the time domain or the frequency domain, which inadequately captures the trends and periodic characteristics. To address this issue, we propose a wavelet learning framework to model complex temporal dependencies of the time series data. The wavelet domain integrates both time and frequency information, allowing for the analysis of local characteristics of signals at different scales. Additionally, the Softmax self-attention mechanism used by Transformers has quadratic complexity, which leads to excessive computational costs when capturing long-term dependencies. Therefore, we propose a novel attention mechanism: Rotary Route Attention (RoRA). Unlike Softmax attention, RoRA utilizes rotary position embeddings to inject relative positional information to sequence tokens and introduces a small number of routing tokens $r$ to aggregate information from the $KV$ matrices and redistribute it to the $Q$ matrix, offering linear complexity. We further propose WaveRoRA, which leverages RoRA to capture inter-series dependencies in the wavelet domain. We conduct extensive experiments on eight real-world datasets. The results indicate that WaveRoRA outperforms existing state-of-the-art models while maintaining lower computational costs. Our code is available at https://github.com/Leopold2333/WaveRoRA.
Multivariate time series forecasting plays a crucial role in various real-world applications. Significant efforts have been made to integrate advanced network architectures and training strategies that enhance the capture of temporal dependencies, thereby improving forecasting accuracy. On the other hand, mainstream approaches typically utilize a single unified model with simplistic channel-mixing embedding or cross-channel attention operations to account for the critical intricate inter-channel dependencies. Moreover, some methods even trade capacity for robust prediction based on the channel-independent assumption. Nonetheless, as time series data may display distinct evolving patterns due to the unique characteristics of each channel (including multiple strong seasonalities and trend changes), the unified modeling methods could yield suboptimal results. To this end, we propose DisenTS, a tailored framework for modeling disentangled channel evolving patterns in general multivariate time series forecasting. The central idea of DisenTS is to model the potential diverse patterns within the multivariate time series data in a decoupled manner. Technically, the framework employs multiple distinct forecasting models, each tasked with uncovering a unique evolving pattern. To guide the learning process without supervision of pattern partition, we introduce a novel Forecaster Aware Gate (FAG) module that generates the routing signals adaptively according to both the forecasters' states and input series' characteristics. The forecasters' states are derived from the Linear Weight Approximation (LWA) strategy, which quantizes the complex deep neural networks into compact matrices. Additionally, the Similarity Constraint (SC) is further proposed to guide each model to specialize in an underlying pattern by minimizing the mutual information between the representations.
World modelling is essential for understanding and predicting the dynamics of complex systems by learning both spatial and temporal dependencies. However, current frameworks, such as Transformers and selective state-space models like Mambas, exhibit limitations in efficiently encoding spatial and temporal structures, particularly in scenarios requiring long-term high-dimensional sequence modelling. To address these issues, we propose a novel recurrent framework, the \textbf{FACT}ored \textbf{S}tate-space (\textbf{FACTS}) model, for spatial-temporal world modelling. The FACTS framework constructs a graph-structured memory with a routing mechanism that learns permutable memory representations, ensuring invariance to input permutations while adapting through selective state-space propagation. Furthermore, FACTS supports parallel computation of high-dimensional sequences. We empirically evaluate FACTS across diverse tasks, including multivariate time series forecasting and object-centric world modelling, demonstrating that it consistently outperforms or matches specialised state-of-the-art models, despite its general-purpose world modelling design.