acceptodds
Under review as a conference paper at ICLR 2027

Learning Continuous and Discrete Dynamics for Multivariate Time Series Anomaly Detection

Abstract

Anomaly detection for multivariate time series plays an important role in many applications, enabling, e.g., risk monitoring in cyber-physical systems. While existing methods achieve good results on continuous variates, they struggle when having to learn both continuous and discrete dynamics across continuous time. Further, existing methods simply sum up reconstruction or contrastive errors from each variate to obtain final anomaly scores. They suffer from variates with different measurement units. Thus, we propose TAD-UP that learns both continuous and discrete dynamics for Time series Anomaly Detection via Unified Probabilistic modeling. First, we propose two co-dependent branches of efficient neural ordinary differential equations with the compound Poisson process to learn continuous and discrete dynamics for different variates. We also propose a gate mechanism to learn correlations among different dynamics. Second, we propose to model a joint probability distribution for anomaly detection. The resulting model is optimized using Maximum Likelihood Estimation on joint variates, instead of using reconstruction or contrastive losses on each variate, which tackles the problem of different measurement units. Experiments on nine real-world datasets from different domains offer evidence that TAD-UP is capable of state-of-the-art performance and better efficiency tradeoff even compared to foundation models.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Anomaly Detection, Continuous and Discrete Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_dEkzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TAD-UP for multivariate time series anomaly detection involving both continuous and discrete variables. The model uses two branches: a continuous branch parameterized by gated temporal convolutional networks within a neural ODE, and a discrete branch incorporating a state-dependent jump process. Cross-branch gates capture dependencies between variable types. Gaussian densities and categorical probabilities are combined to construct the training objective and anomaly scores.

Critical strengths

Continuous measurements and categorical states have different semantics and output spaces. Using separate embeddings, dynamics branches, and decoders while exchanging information through gates provides a plausible way to capture these differences and their dependencies. Experimental results demonstrate the effectiveness of TAD-UP.

Critical weaknesses
  1. The technical contribution appears incremental. TAD-UP primarily combines established components: neural ODEs, neural jump SDEs, and gated TCNs. In particular, learning continuous flows together with state-dependent jumps has already been explored by Jia and Benson (2019).

  2. The Poisson jump assumption is insufficiently justified. The specific waiting-time assumptions of the Poisson jump process may not reflect the actual timing of observed events.

Questions for authors
  • Could you add an ablation study that replaces the jump SDE with an ordinary ODE while keeping everything else unchanged? This would show whether the jump mechanism itself improves performance.

  • Could you explain why TAD-UP achieves inference latency comparable to the baselines despite the additional cost of numerically solving its continuous and jump dynamics? Please specify the solver, step size, and number of function evaluations.

  • Equations (9) and (11) omit the exponent NCN_C on 2π2\pi under the square root.

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

This paper proposes TAD-UP for unsupervised multivariate time series anomaly detection involving both continuous-valued and discrete-valued variables. The method uses two co-dependent neural differential-equation branches: a continuous co-ODE for continuous temporal dynamics and a discrete co-ODE incorporating a compound Poisson process to represent jumps. Gated temporal convolutional networks (TCNs) introduce cross-branch conditioning so that the two types of dynamics can inform each other.

For anomaly scoring, the authors propose a unified probabilistic formulation. The continuous observations are modeled using Gaussian negative log-likelihood terms, including a decoder-based covariance and an empirical covariance term, while discrete observations are modeled using categorical softmax probabilities and cross-entropy. The resulting losses are combined, and low observation likelihood is used to identify anomalous timestamps. Experiments across nine real-world datasets and component ablations report competitive anomaly-detection performance and an accuracy–latency trade-off.

Critical strengths

1. Meaningful integration of heterogeneous temporal dynamics and probabilistic anomaly scoring.

The paper addresses the practically relevant setting in which continuous measurements and discrete system states coexist and may be correlated. Its integration of dedicated dynamic branches, cross-branch gated interactions, and probabilistic anomaly scoring offers a coherent way to model heterogeneous temporal patterns rather than treating all variables identically. This overall system design is a meaningful aspect of the work, even though several constituent modeling ideas have precedents.

Critical weaknesses

1. Insufficient differentiation from closely related prior work.

Although the paper evaluates numerous baselines, it does not clearly distinguish TAD-UP from prior work on neural jump dynamics (Jia and Benson, 2019), probabilistic anomaly scoring (DAGMM and OmniAnomaly), and mixed-type anomaly detection (ADMNC). While these precedents do not invalidate the proposed approach, the specific methodological advances of the co-dependent branches, cross-branch gating, and unified probabilistic modeling remain insufficiently explained, making the incremental novelty difficult to assess.

2. Insufficient mathematical justification of the unified probabilistic formulation.

In Section 3.5, Eqs. (9) and (11) define two Gaussian negative log-likelihood terms for the same continuous observation, and the proposed loss simply adds these two terms. However, the product of two Gaussian densities is generally not normalized, and the required normalization factor may depend on the model parameters. Therefore, the proposed objective is not clearly justified as maximum likelihood estimation of a normalized joint distribution. Additionally, the Gaussian normalization constants appear to omit the dimension-dependent factor. These issues require clarification, as unified probabilistic modeling is a central contribution of the paper.

Questions for authors

1. Distinction from prior work. Could the authors clarify what is technically new in the co-dependent dynamic branches and cross-branch gating relative to earlier continuous–discrete or jump-dynamics models, and how the proposed probabilistic anomaly scoring differs from established likelihood-based anomaly detection and mixed-type probabilistic modeling? I am not asking for novelty of every individual component, but for a more precise account of the contribution of their combination.

2. Probabilistic validity of the objective. What is the explicitly normalized joint probability distribution corresponding to Eqs. (9)–(13)? In particular, since Eqs. (9) and (11) score the same continuous observation using two Gaussian terms, how is their product normalized, and is the resulting normalizer included in optimization and inference? If the sum is instead intended as a composite likelihood or energy-based loss, could the authors revise the MLE/joint-probability claims accordingly? Please also clarify the Gaussian normalizing factors in Eqs. (9) and (11).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.