EVOLVE: Uncertainty-Aware Multivariate Time Series Anomaly Prediction with Adaptive Prototype Memories

Jianxin Jin; Kai Zhao; Shiyan Hu; Yang Shu; Peng Chen; Meng Wang; Zhongwen Rao; Yijie Li

cs.LGOct 23, 2026

EVOLVE: Uncertainty-Aware Multivariate Time Series Anomaly Prediction with Adaptive Prototype Memories

Authors: Jianxin Jin, Kai Zhao, Shiyan Hu, Yang Shu, Peng Chen, Meng Wang, Zhongwen Rao, Yijie Li,

Organizations: East China Normal University · Aalborg University · APPLab, Huawei · Huawei Technologies Co., Ltd ·

Abstract

Unsupervised multivariate time series anomaly prediction (AP) aims to estimate future anomaly risks from current observations, providing early-warning signals for proactive intervention and maintenance in complex systems. Existing unsupervised AP methods follow a perturbation-based paradigm, where predefined or random perturbations are injected into historical windows to simulate abnormal future evolution. However, these methods share perturbations across different variables and contexts, and learn deterministic history-to-future evolution. In reality, anomaly evolution is variable- and context-dependent: the same perturbation may be amplified into abnormal future dynamics in one variable or context, while being absorbed as normal fluctuation in another. In addition, future evolution is not uniquely determined by a historical window, since similar histories may lead to multiple plausible trajectories, such as gradual degradation or abrupt changes. To address these problems, we propose EVOLVE, a unified framework that models unsupervised AP via variable- and context-conditioned perturbations and uncertainty-aware future reasoning. EVOLVE constructs context-specific anomaly perturbations through an adaptive anomaly prototype memory, which can dynamically update and retrieve relevant perturbations at the variable level. EVOLVE models uncertainty by sampling multiple latent historical states and decoding through complementary smooth and deviation-amplifying evolution hypotheses, and the disagreement among decoded future trajectories is used to estimate future anomaly risk. Extensive experiments on six real-world benchmark datasets demonstrate that EVOLVE consistently outperforms state-of-the-art anomaly prediction baselines.

Figures & tables

Fig. 1: Different impacts of similar historical perturbations on future evolution under time series heterogeneity. (a) Perturbations are amplified with abnormal evolution. (b) Perturbations are absorbed with normal evolution.

Fig. 2: Overall architecture of EVOLVE. Given a historical window, EVOLVE first encodes the input into a normal historical representation $H_{T}^{n}$ . During training, Adaptive Anomaly Injection (AAI) retrieves variable-conditioned perturbations from the Anomaly Prototype Memory (APM) and constructs an anomaly-aware historical representation $H_{T}^{a}$ , while Anomaly Pattern Analysis (APA) progressively updates the dynamic prototype bank using residual-based anomaly-aware future representations. Both the normal and anomaly-aware historical views are then processed by Uncertainty-Aware Latent Modeling (UALM), which samples multiple latent semantic hypotheses and estimates their evolution-mode probabilities. The Dual-Structured Decoder (DSD) decodes these hypotheses through smooth and deviation-amplifying evolution hypotheses to generate multiple plausible future trajectories. During inference, anomaly-aware injection, APA, and dynamic prototype updating are disabled, and future anomaly scores are computed from the predictive disagreement among decoded trajectories.

Dataset	MSL			PSM			SMD			SWAN			GECCO			Genisis
Metric	Aff-F1	V-ROC	V-PR	Aff-F1	V-ROC	V-PR	Aff-F1	V-ROC	V-PR	Aff-F1	V-ROC	V-PR	Aff-F1	V-ROC	V-PR	Aff-F1	V-ROC	V-PR
LODA	63.78	53.53	10.57	61.02	48.28	30.41	57.93	39.14	5.34	37.11	18.38	30.04	39.04	43.27	2.23	54.25	46.62	0.93
IForest	66.59	53.31	9.17	65.25	50.17	31.86	58.69	39.91	3.53	30.00	27.45	32.54	46.49	39.86	1.93	61.25	50.06	1.41
TranAD	65.72	52.90	16.77	64.01	43.13	34.31	63.43	42.55	4.53	51.30	37.07	30.44	41.13	22.76	1.66	63.26	64.33	0.94
DCdetector	59.41	49.68	10.69	66.08	37.49	27.02	65.51	40.43	4.69	52.66	32.92	32.99	52.62	41.21	1.71	67.56	53.82	1.25
ATransformer	54.57	48.66	13.81	53.20	43.56	30.09	47.70	47.28	4.08	40.99	27.51	30.84	49.58	43.54	0.84	45.26	42.20	0.35

TABLE I: Anomaly prediction performance of all models on six datasets

Dataset		MSL			SMD			SWAN
Smooth	Alert	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR
✗	✗	70.21	65.41	20.88	72.65	54.65	6.85	63.40	40.02	35.96
✓	✗	68.75	64.56	20.55	70.97	56.40	6.76	63.36	39.26	35.16
✗	✓	68.78	63.59	18.55	72.05	55.97	6.73	61.10	42.08	36.52
✓	✓	72.70	70.23	23.81	72.88	57.77	6.93	65.08	42.47	37.81

TABLE II: ABLATION OF THE DUAL-STRUCTURED DECODER.

Dataset	MSL			SMD			SWAN
Metric	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR
w/o Dynamically Adaptive Anomalies	70.51	64.63	21.24	70.57	51.31	6.26	59.50	41.58	36.55
w/o Anomaly Pattern Analysis	70.57	65.99	20.98	71.54	56.85	6.62	60.21	40.62	36.06
w/o Adaptive Latent Modeling	69.55	66.09	21.55	71.07	56.03	6.65	61.92	42.12	36.51
w/o Adversarial Anomaly Generation	70.21	63.35	20.84	70.78	57.06	6.50	62.68	39.94	36.06
EVOLVE	72.70	70.23	23.81	72.88	57.77	6.93	65.08	42.47	37.81

TABLE III: ABLATION OF KEY COMPONENTS.

Dataset	MSL			SMD			SWAN
Metric	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR	AFF-F1	V-ROC	V-PR
w/o MSE Loss	70.44	65.66	21.01	71.96	55.68	6.61	64.88	40.47	35.67
w/o VAR Loss	69.54	64.51	20.17	70.90	56.23	6.36	60.80	41.98	35.24
w/o Weights Loss	69.03	66.05	20.11	70.55	54.45	6.19	62.52	41.25	36.60
EVOLVE	72.70	70.23	23.81	72.88	57.77	6.93	65.08	42.47	37.81

TABLE IV: ABLATION OF TRAINING LOSSES.

Fig. 3: Parameter sensitivity analysis of EVOLVE on MSL, SMD, and SWAN using VUS-PR.

Method	VUS-ROC	Training Time	Inference Time	Memory Used
IGCL	66.74	14.09s	0.19s	34.12MB
PAD	60.79	99.20s	0.86s	281.13MB
A2P	62.01	34.87s	1.36s	559.77MB
RED-F	54.96	78.54s	2.15s	567.22MB
TAP	71.04	97.33s	1.85s	496.63MB
FCM	57.04	75.56s	1.97s	443.53MB

TABLE V: Efficiency and effectiveness comparison on MSL.

$N_{\mathrm{mem}}$	Training Time	Memory Used	VUS-ROC
50	65.57s	135MB	75.64
100	67.21s	146MB	76.17
200	71.56s	160MB	77.81
300	73.43s	176MB	76.10
400	76.20s	205MB	75.54
500	77.99s	229MB	74.78

TABLE VI: Scalability analysis with respect to the anomaly prototype memory size

N_{\mathrm{mem}}

Fig. 4: Visualization of predictive anomaly scores on the SMD dataset.

Time-Series Anomaly Detection

Time-Series Anomaly Detection identifies unusual patterns or events in sequential data, from sensor readings and network traffic to health and industrial telemetry.

See all 208 papers

Oct 18, 2025

Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection

Real-world multivariate time series anomalies are rare and often unlabeled. Additionally, prevailing methods rely on increasingly complex architectures tuned to benchmarks, detecting only fragments of anomalous segments and overstating performance. In this paper, we introduce OracleAD, a simple and interpretable unsupervised framework for multivariate time series anomaly detection. OracleAD encodes each variable's past sequence into a single causal embedding to jointly predict the present time point and reconstruct the input window, effectively modeling temporal dynamics. These embeddings then undergo a self-attention mechanism to project them into a shared latent space and capture spatial relationships. These relationships are not static, since they are modeled by a property that emerges from each variable's temporal dynamics. The projected embeddings are aligned to a Stable Latent Structure (SLS) representing normal-state relationships. Anomalies are identified using a dual scoring mechanism based on prediction error and deviation from the SLS, enabling fine-grained anomaly diagnosis at each time point and across individual variables. Since any noticeable SLS deviation originates from embeddings that violate the learned temporal causality of normal data, OracleAD directly pinpoints the root-cause variables at the embedding level. OracleAD achieves state-of-the-art results across multiple real-world datasets and evaluation protocols, while remaining interpretable through SLS.

Dongchan Cho, Jiho Han, Keumyeong Kang +3

Industrial AI Lab, SimPlatform Co. Ltd. Affiliate Research Institute

Time-Series Anomaly Detection Unsupervised Anomaly Detection Interpretable Anomaly Detection Anomaly Localization Reconstruction-Based Anomaly Detection

Ask PDF

Feb 19, 2026

Forecasting Anomaly Precursors via Uncertainty-Aware Time-Series Ensembles

Detecting anomalies in time-series data is critical in domains such as industrial operations, finance, and cybersecurity, where early identification of abnormal patterns is essential for ensuring system reliability and enabling preventive maintenance. However, most existing methods are reactive: they detect anomalies only after they occur and lack the capability to provide proactive early warning signals. In this paper, we propose FATE (Forecasting Anomalies with Time-series Ensembles), a novel unsupervised framework for detecting Precursors-of-Anomaly (PoA) by quantifying predictive uncertainty from a diverse ensemble of time-series forecasting models. Unlike prior approaches that rely on reconstruction errors or require ground-truth labels, FATE anticipates future values and leverages ensemble disagreement to signal early signs of potential anomalies without access to target values at inference time. To rigorously evaluate PoA detection, we introduce Precursor Time-series Aware Precision and Recall (PTaPR), a new metric that extends the traditional Time-series Aware Precision and Recall (TaPR) by jointly assessing segment-level accuracy, within-segment coverage, and temporal promptness of early predictions. This enables a more holistic assessment of early warning capabilities that existing metrics overlook. Experiments on five real-world benchmark datasets show that FATE achieves an average improvement of 19.9 percentage points in PTaPR AUC and 20.02 percentage points in early detection F1 score, outperforming baselines while requiring no anomaly labels. These results demonstrate the effectiveness and practicality of FATE for real-time unsupervised early warning in complex time-series environments.

Hyeongwon Kang, Jinwoo Park, Seunghun Han +1

Department of Industrial & Management Engineering, Korea University, 126-16 Anam-dong 5-ga, Seongbuk-gu, Seoul, Republic of Korea · Department of Industrial Engineering, Seoul National University, Gwanak-ro 1, Gwanak-gu, Seoul, Republic of Korea · LG CNS, 71, Magokjungang 8-ro, Gangseo-gu, Seoul, Republic of Korea

Early Warning Detection Time-Series Anomaly Detection Unsupervised Anomaly Detection

Ask PDF

Aug 12, 2026

JAPE: Joint Anomaly Prediction and Intrinsic Explanation in Multivariate Time Series

Multivariate time-series anomaly prediction aims to identify whether and when anomalies will occur over a future horizon from historical observations. Existing methods primarily characterize anomalies as deviations in future numerical values, which may overlook subtle dependency changes induced by weak anomaly precursors and provide no native variable-level explanation together with the alert. To bridge these gaps, we propose JAPE, a Joint Anomaly Prediction and Explanation framework that lifts anomaly prediction from numerical-deviation modeling to dependency-structure modeling. JAPE is the first anomaly prediction framework to explicitly model evolving dependency structures for both point-wise alerting and native variable-level explanation. Specifically, JAPE (i) proposes a Decoupled Spatio-Temporal Representation (DSTR) backbone that decouples temporal and spatial modeling and captures lag-aware dependencies via learnable lag aggregation, thereby perceiving structural precursors before numerical deviations emerge; (ii) designs a dual-view alerting mechanism that fuses numerical forecasts with evolving dependency graphs for point-wise anomaly prediction, capturing structural evidence even under subtle numerical deviations; and (iii) presents Native Predictive Explanation (NPE), which directly reuses the predicted dependency graphs to rank variables by structural deviations without additional models or training. Extensive experiments on five real-world benchmarks across three prediction horizons demonstrate that JAPE improves average F1 and AUC-PR by 19.7% and 41.3%, respectively, while improving explainability with 26.6% gain in MRR.

Yian Wei, Yuanyuan Yao, Lu Chen +2

Zhejiang University · RMIT University · Aalborg University

Interpretable Anomaly Detection Time-Series Anomaly Detection Anomaly Detection

Ask PDF

Abstract

Time-Series Anomaly Detection

Explore similar work

Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection

Forecasting Anomaly Precursors via Uncertainty-Aware Time-Series Ensembles

JAPE: Joint Anomaly Prediction and Intrinsic Explanation in Multivariate Time Series