Tensorial Change Analysis Using Probabilistic Tensor Regression

Tsuyoshi IdéAAAI 2019 · DOI

Introduction

In complex industrial systems, such as semiconductor manufacturing, data is inherently multi-dimensional. Detecting when a system has changed is a common task, but identifying why it changed—which specific component or time step is responsible—remains a significant challenge. This paper, presented at AAAI-19, introduces Bayesian Alternating Least Squares (BALS), a framework for supervised tensorial change analysis that provides both high-accuracy detection and interpretable "change scores" for root-cause diagnosis.

The Problem of Tensorial Change

Traditional change detection often focuses on identifying outliers in raw data. However, industrial processes require supervised change detection: identifying shifts in the relationship between multi-dimensional inputs (e.g., sensor readings over multiple etching steps) and a scalar output (e.g., product quality).

Conceptual diagram of tensorial data $\mathcal{X}$ and performance metric $y$ over time.

Existing approaches typically suffer from two main issues:

  1. Vectorization: Flattening tensors into vectors destroys the structural relationships between dimensions, such as the link between a specific sensor and a specific process step.
  2. Computational Complexity: Bayesian tensor models often rely on slow Monte Carlo sampling, making them impractical for real-time monitoring.

The BALS Framework

The proposed BALS method addresses these challenges by combining a probabilistic linear model with Canonical Polyadic (CP) decomposition. This decomposition breaks the large coefficient tensor into a sum of rank-1 components, preserving the multi-way structure of the data.

To ensure efficiency, the authors employ Variational Expectation-Maximization (vEM). This provides analytic, closed-form update equations that are significantly faster than sampling-based methods. Most importantly, BALS enables "Change Analysis Scores" by calculating the Kullback-Leibler (KL) divergence between the posterior distribution of a "golden" (normal) period and a test period. This allows engineers to pinpoint exactly which variables and time steps contributed to a detected anomaly.

Comparative Performance

The authors evaluated BALS against state-of-the-art Gaussian Process Regression (GPR) variants across synthetic and real-world datasets.

Comparison of RMSE and total computation time for BALS vs. n-GPR and d-GPR.

As shown in the experimental results, BALS achieves near-zero Root Mean Square Error (RMSE) on synthetic benchmarks while maintaining high computational efficiency. In synthetic tests, BALS was nearly 20x faster than mode-wise kernel methods (d-GPR). On the London School dataset, BALS achieved an AUC of 0.96, significantly outperforming GPR variants (0.85–0.88).

Real-World Utility: Semiconductor Etching

In a case study involving semiconductor etching data (35 sensors over 14 steps), BALS successfully identified a specific sensor ($x_{12}$) at a specific step (step 3) as the root cause of a quality excursion. This diagnosis was later confirmed by raw data plots showing increased fluctuation in that specific sensor, demonstrating the model's ability to translate abstract statistical shifts into actionable industrial insights.

Conclusion

BALS provides a robust, efficient, and interpretable solution for tensorial change analysis. By respecting the natural tensor structure of industrial data and providing information-theoretic diagnosis scores, it bridges the gap between high-performance machine learning and practical root-cause analysis. Future work may involve extending the model to handle non-linear relationships and automating the selection of the CP rank.

Summary updated

BibTeX
@inproceedings{ide2019tensorial,
  author = {Tsuyoshi Idé},
  title = {{Tensorial Change Analysis Using Probabilistic Tensor Regression}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33013902},
}
Figure