🤖 machine learning

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

JAPE is a novel framework for multivariate time-series anomaly prediction that shifts from modeling numerical deviations to capturing evolving dependency structures, enabling both early detection of subtle precursors and native variable-level explanations without additional training.

Original authors: Yian Wei, Yuanyuan Yao, Lu Chen, Xiangmin Zhou, Tianyi Li

Published 2026-08-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yian Wei, Yuanyuan Yao, Lu Chen, Xiangmin Zhou, Tianyi Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a massive, high-tech spaceship. Your ship isn't just one big engine; it's a complex web of thousands of sensors, pipes, and computers all talking to each other. Most of the time, they hum along in perfect harmony. But sometimes, things go wrong. In the world of data science, this is called "multivariate time series"—a fancy way of saying "lots of different numbers changing over time." For years, the standard way to catch trouble was to wait until a sensor started screaming. If the temperature in the engine room suddenly jumped from 100 to 200 degrees, an alarm would blare. But by then, the engine might already be melting. It's like waiting for the smoke to fill the room before you realize there's a fire.

Scientists have been trying to build better alarms that can predict a disaster before it happens. They look at the history of the numbers to guess what's coming next. However, there's a catch. Sometimes, the numbers don't change much at all before a crash. Instead, the relationships between the sensors change. It's like two friends who usually walk side-by-side suddenly starting to walk in opposite directions. The numbers (their speed) might look normal, but their connection is broken. Existing methods often miss these subtle "friendship breakups" because they are too focused on the numbers themselves. They wait for the smoke, missing the moment the friends stopped talking.

This is where a new framework called JAPE (Joint Anomaly Prediction and Intrinsic Explanation) comes in. Think of JAPE as a super-smart detective who doesn't just watch the numbers; they watch the relationships between the sensors. The researchers, led by Yian Wei and colleagues from Zhejiang University and others, propose that to predict a disaster, you need to understand how the system's "social network" is evolving.

Here is how JAPE works, broken down into its three superpowers:

1. The "Two-Track" Detective (DSTR)
Most old methods try to learn how a single sensor changes over time and how sensors talk to each other all at once, like trying to listen to a whole choir while also trying to hear one singer. JAPE separates these tasks. It has a "Time Track" that watches how each sensor changes on its own, and a "Space Track" that watches how the sensors influence each other. Crucially, the Space Track is "lag-aware." It understands that Sensor A might influence Sensor B a few seconds later, not instantly. By separating these tracks, JAPE can spot a tiny, subtle shift in how sensors relate to each other before the numbers themselves start to go crazy. It's like noticing that two friends are whispering nervously to each other before they even start arguing.

2. The "Double-Check" Alarm (Dual-View Alerting)
Once JAPE has its two tracks running, it doesn't just pick one to decide if an alarm should go off. It uses a "Dual-View" system. One view looks at the predicted numbers (the "what will happen" view), and the other looks at the predicted relationships (the "how they will connect" view). It fuses these two perspectives together. Even if the numbers look perfectly normal, if the relationship map shows a weird, sudden break in the pattern, JAPE sounds the alarm. This allows it to catch "weak" precursors that other methods miss. In tests on five real-world datasets (like data from servers and machines), this approach helped JAPE improve its ability to find anomalies by 19.7% in accuracy and 41.3% in ranking quality compared to the best existing methods.

3. The "Instant Explanation" (NPE)
Usually, when an alarm goes off, a human has to spend hours digging through logs to figure out why it happened and which sensor started the trouble. JAPE skips this step. Because it built a map of the relationships while making the prediction, it can instantly point to the "culprit." It uses a score called the "Graph Deviation Score" to rank which variables are acting strangely compared to their normal behavior. It's like the detective not only saying "Fire!" but immediately pointing at the specific candle that was knocked over. The paper shows that this method improves the ability to find the right variable by 26.6% compared to other methods, and it does this without needing any extra training or a second look at the data.

The researchers tested JAPE on five different real-world benchmarks, including data from server clusters and industrial machines. They found that JAPE didn't just find more problems; it found them earlier and explained them better. For example, on one dataset, it improved the ability to rank the correct "bad" variables from a 23.5% success rate to 33.3%. The best part? It does all this without slowing things down. The "explanation" part takes almost no extra time (less than 0.5% of the total runtime) because it reuses the work it already did to make the prediction.

In short, JAPE suggests that to predict the future of complex systems, we shouldn't just watch the numbers; we should watch how the numbers talk to each other. By treating the system like a living social network rather than a list of statistics, JAPE offers a way to catch trouble before it starts and tell you exactly who to blame, all in a single, speedy glance.

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

1. Problem Definition

Multivariate time-series (MTS) anomaly prediction aims to anticipate whether and when anomalies will occur within a future horizon based on historical observations. Unlike traditional anomaly detection, which reacts to faults after they manifest, anomaly prediction requires proactive identification of onset times.

The paper identifies three fundamental limitations in existing value-centric approaches (which model anomalies primarily as deviations in future numerical values):

Delayed Detection (L1): Weak anomaly precursors, often manifesting as subtle changes in inter-variable dependencies rather than numerical spikes, are frequently absorbed as normal fluctuations by models optimized for numerical reconstruction. This causes alerts to lag behind fault progression.
Limited Anomaly Discriminability (L2): During anomaly propagation, source variables and affected variables often exhibit similar numerical deviations. Value-centric methods struggle to distinguish the root cause from propagated responses, especially when source changes are subtle.
Lack of Intrinsic Explainability (L3): Existing methods detect anomalies but cannot natively identify the specific variables involved. They require separate, post-hoc analysis pipelines for interpretation, which adds computational overhead and delays root-cause diagnosis.

2. Methodology: The JAPE Framework

JAPE (Joint Anomaly Prediction and Explanation) shifts the paradigm from numerical-deviation modeling to dependency-structure modeling. It is trained in a setting where a small number of anomaly labels are available. The framework consists of three core components:

2.1 Decoupled Spatio-Temporal Representation (DSTR)

To address the masking of weak structural precursors by dominant temporal patterns, JAPE decouples temporal evolution from inter-variable dependency modeling into two dedicated axes within a backbone architecture:

Temporal Axis: Models the evolution of individual variables independently using channel-independent temporal modeling (similar to PatchTST), preserving the temporal dynamics of each series.
Spatial Axis: Explicitly models inter-variable dependencies using a Lag-Aware Directed Graph Learning mechanism.
- It employs a learnable lag aggregation mechanism to construct lagged representations of variables, capturing delayed influences.
- It uses a directional contrast function to compute asymmetric dependency scores (e.g., how the past of variable $v$ predicts the present of variable $u$ ), distinguishing directionality.
- This axis generates a sequence of dynamic directed dependency graphs that evolve over time, preserving structural changes before they become numerically apparent.
Future-Oriented Branch: Extends the spatial dependency modeling into the prediction horizon, constructing future dependency graphs to provide a "before-and-after" perspective for alerting.

2.2 Dual-View Alerting Mechanism

This component fuses two complementary views to estimate point-wise anomaly probabilities:

Numerical View: Derived from the predicted future time series ( $\hat{X}$ ), capturing observable responses of individual variables.
Structural View: Derived from the evolving dependency graphs ( $A$ ) produced by DSTR. To handle variable input dimensions, the graphs are summarized into compact structural descriptors (13-dimensional vectors) capturing statistics like graph energy, entropy, and dependency shifts.
Fusion: A cross-attention mechanism adaptively fuses the numerical tokens and structural tokens. This allows the model to exploit source-aware structural evidence even when numerical deviations are subtle, significantly improving anomaly discrimination.

2.3 Native Predictive Explanation (NPE)

NPE provides variable-level explanations without additional models or training:

Mechanism: It reuses the dependency graphs generated during the prediction phase.
Graph Deviation Score (GDS): It calculates a deviation score by comparing the predicted dependency graph against a normal reference graph. The score combines:
- Direct Deviation: Immediate changes in outgoing and incoming dependencies.
- Multi-hop Deviation: Propagation of structural changes through dependency chains (up to $K$ hops).
Output: Variables are ranked based on their GDS. Since the graphs are already constructed during inference, this process incurs negligible computational cost and eliminates the need for post-hoc diagnostic pipelines.

3. Key Contributions

DSTR Backbone: The first anomaly prediction framework to explicitly decouple temporal evolution from dependency modeling, preserving weak structural precursors via dynamic directed graphs.
Dual-View Alerting: A mechanism that fuses numerical forecasts with dynamic dependency structures, enabling robust detection even when numerical signals are weak.
Native Predictive Explanation (NPE): A method that generates variable-level rankings at zero additional cost by reusing predicted dependency graphs, removing the need for separate attribution models.
Empirical Validation: Extensive experiments demonstrating significant improvements over state-of-the-art baselines.

4. Experimental Results

The authors evaluated JAPE on five real-world benchmarks (SMD, WADI, MSL, PSM, EXATHLON) across three prediction horizons (50, 100, 200 time steps).

Prediction Performance: JAPE achieved an average F1 score of 62.9% and an AUC-PR of 70.9%. This represents a 19.7% improvement in F1 and a 41.3% improvement in AUC-PR over the strongest baseline (A2P-Sup). The gains were particularly pronounced on datasets with complex dependency evolution (e.g., MSL), where numerical-only methods failed to capture weak precursors.
Explainability: On datasets with ground-truth variable annotations (SMD, WADI), JAPE improved the Mean Reciprocal Rank (MRR) by 26.6% compared to the best baseline, demonstrating superior ability to identify the true anomaly source variables.
Efficiency: JAPE demonstrated high training efficiency, reducing per-epoch training time by up to 70.5% compared to A2P on the SMD dataset. The NPE component accounted for less than 0.5% of the total runtime.

5. Significance and Claims

The paper claims that JAPE is the first anomaly prediction framework to explicitly model evolving dependency structures to jointly achieve early point-wise alerting and native variable-level explanation.

The authors argue that by lifting anomaly prediction from numerical-deviation modeling to dependency-structure modeling, JAPE addresses the core limitations of existing methods:

It detects anomalies earlier by perceiving structural precursors before numerical deviations emerge.
It improves discriminability by distinguishing source variables from propagated responses through structural evidence.
It provides intrinsic explainability without the latency or cost of post-hoc analysis.

The work suggests that inter-variable relationships provide meaningful structural evidence for system changes, and explicitly modeling these evolving structures is essential for proactive monitoring in complex distributed systems. Future work is proposed to scale this approach to larger systems and explore adaptive graph sparsification.