Capturing Pattern Relationships in Multivariate Time Series through an Episode-Cluster GCN

Minjeong Oh*, Jaesoo Yoo**, Dojin Choi* | *Changwon National University, **Chungbuk National University, Korea

omiin.jeong@gmail.com, yjs@chungbuk.ac.kr, dojinchoi@changwon.ac.kr

1

Introduction

Research Objective

Improving multivariate time-series forecasting accuracy

Sensor Graph

Existing Approaches

  • GCN-based models rely mainly on inter-channel correlations
  • They treat sensors (channels) as fixed graph nodes
  • Uniform segmentation fails to capture meaningful pattern boundaries
Episode Graph

Proposed Approach

  • Detect meaningful change points using surprisal-based episode segmentation
  • Improve forecasting accuracy by capturing intra-channel pattern structures
2

Background

  • Introduce the concept of surprisal-based segmentation for detecting meaningful pattern changes
  • In EM-LLM, surprisal is computed using the model's prediction errors to identify unexpected changes

    • - Surprisal quantification: Negative Log-Likelihood

  • We redefine this surprisal measure using a Gaussian Negative Log-Likelihood (NLL) for time-series data
-log P(xt | x1,...,xt-1; θ) > T    with    T = μt-τ:t + γσt-τ:t
Surprise Timeline
3

Methodology

1 Use Informer to predict mean and standard deviation
2 Compute surprisal using Gaussian NLL based on the predicted mean and standard deviation
3 Detect boundaries via surprisal thresholding
4 Segment into episodes and assign cluster indices
5 Build an episode-level node graph
6 Apply cluster-level Graph Convolution
7 Form the final Episode-Cluster GCN representation
Pipeline
4

Datasets

  • Experiments are conducted on widely used benchmark datasets for multivariate time-series forecasting
  • To verify the robustness of episode-boundary detection on periodic patterns, we select two strongly periodic datasets: ETTm1 and Weather
ETTm1 Weather
Desc. Electricity Transformer Temperature dataset Meteorological time-series
with strong seasonal variations
(temperature, humidity, wind speed, pressure, etc.)
Sampling 15 min 10 min
# of features 7 21
5

Experiments

Backbone Model

  • Model: Informer
  • Objective: Predict mean & standard deviation
  • Loss: Gaussian Negative Log-Likelihood (NLL)

Forecasting Model

  • Lookback window: 512 time steps
  • Forecasting horizon: 92 time steps
  • Evaluation metric: MSE(Mean Squared Error)

Forecasting Results (MSE)

GTA Episode-Cluster GCN (Our)
ETTm1 0.0106 0.0098
Weather 0.0084 0.0059
6

Conclusions

Contributions

  • Our method captures natural pattern changes through episode boundary detection
  • It enhances structural graph representations
  • It achieves better forecasting accuracy and improves interpretability

Future Works

  • We will evaluate the model with various lookback and horizon settings
  • We will explore alternative backbone models
Sensor Visualization