Research map / Time series / Anomaly detection
Channel Matters: Estimating Channel Influence for Multivariate Time Series
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Channel-wise Influence (ChInf), a TracIn-based per-channel influence matrix for multivariate time series, and applies it to unsupervised MTS anomaly detection and channel pruning for forecasting.
Problem
Existing influence functions treat a whole multivariate time series sample as one unit and cannot distinguish the influence of individual channels, so they perform poorly on MTS anomaly detection and data pruning.
Approach
Decompose the TracIn first-order influence into a sum over channel pairs, yielding a channel-wise influence matrix whose diagonal (self-influence per channel) serves as an anomaly score and whose ranking enables selecting representative channels for pruning.
Key results
- On SWaT, the ChInf-based iTransformer reaches F1 84.0 versus 83.7 for the original Inverted Transformer; on SMAP, ChInf with GCN-LSTM reaches F1 48.0 versus 42.6 for the base model.
- For channel pruning on ECL, iTransformer with influence-based selection matches full-variate MSE (0.150 vs 0.148) while retaining 50% of channels.
- Self-influence based anomaly detection is reported to be at millisecond per-timestep latency (1.4ms on SWaT with GCN-LSTM).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ChInf instead of the base anomaly detector's original scoring method | Inverted Transformer · SWaT | F1: 83.7 -> 84.0 | Helped |
| Apply ChInf instead of the base anomaly detector's original scoring method | Inverted Transformer · SMD | F1: 55.9 -> 59.1 | Helped |
| Apply ChInf instead of the base anomaly detector's original scoring method | 1-Layer GCN-LSTM · SMAP | F1: 42.6 -> 48.0 | Helped |
| Apply ChInf instead of the base anomaly detector's original scoring method | 1-Layer GCN-LSTM · MSL | F1: 46.3 -> 47.1 | Helped |
| Apply ChInf instead of the base anomaly detector's original scoring method | Inverted Transformer · WADI | F1: 48.8 -> 50.5 | Helped |
| Replace reconstruction error with ChInf (ablation) | Single Transformer block · SMD | F1: 48.9 -> 52.1 | Helped |
| Replace reconstruction error with ChInf (ablation) | Single block MLPMixer · SMAP | F1: 36.3 -> 48.0 | Helped |
| Use equidistant sampling from the ChInf ranking instead of selecting the most self-influential channels (ablation) | iTransformer · Electricity, 5% channels retained | MSE: 0.360 -> 0.187 | Helped |
| Select channels using ChInf rather than continuous channel selection | iTransformer · ETTh1, 3 channels retained | MSE: 0.502 -> 0.403 | Helped |
| Use channel pruning rather than sample-wise pruning | MTS forecasting · Electricity, Solar Energy, and Traffic | MSE | Helped |
| Use original TracIn rather than ChInf for anomaly detection (ablation) | MTS anomaly detection · SMAP and SMD | F1 | Hurt |
| Increase the number of model parameters used to calculate influence (ablation) | GCN-LSTM anomaly detection · Different anomaly detection datasets | F1 | No clear effect |
Surprising findings
- MoSo-based sample pruning performed worse than random sample pruning on time series forecasting, which the authors attribute to sample dependencies.
- Original TracIn-based anomaly scores were reported to be worse than reconstruction error, while channel-wise ChInf improved over reconstruction error.
Limitations
- Empirical analysis is limited to anomaly detection and channel pruning; other MTS tasks are unexplored.
- Gradients are computed on a subset of parameters (last layer) for efficiency, and results are reported as insensitive to this choice only for GCN-LSTM.
- ChInf may be less effective for globally synchronized anomalies that affect all channels uniformly.
Experimental setup
Datasets: SWaT WADI SMD SMAP MSL Electricity Solar-Energy Traffic ETTh1 ETTm1
Benchmarks: SWaT SMD SMAP MSL WADI Electricity Solar-Energy Traffic
Metrics: F1 precision recall MSE
Baselines: DAGMM OmniAnomaly USAD GDN TranAD AnomalyTransformer PCA ERROR TimeInf LIME selection SHAP selection NFS selection MoSo
Builds on: influence functions TracIn TimeInf iTransformer
Closest papers by meaning
- TimeInf: Time Series Data Contribution via Influence Functions (ICLR 2025 · similarity 0.87)
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.86)
- Rethinking Channel Dependence for Multivariate Time Series Forecasting: Learning from Leading Indicators (ICLR 2024 · similarity 0.86)
- CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching (ICLR 2025 · similarity 0.86)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.85)
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.