Positive Unlabeled Leaning for Time Series Classification

Minh Nhut Nguyen, Xiaoli Li, See-Kiong Ng

Introduction

Time series classification is a critical task in domains ranging from medical diagnostics to industrial monitoring. However, traditional supervised learning requires extensive labeled datasets for both positive and negative classes. In many real-world scenarios, negative examples are entirely missing, and positive examples are extremely scarce—a challenge addressed by Positive-Unlabeled (PU) learning. While PU learning has seen success in text classification, its application to time series is hindered by high feature correlation and noise. Existing methods often fail to find accurate decision boundaries, resulting in high precision but poor recall.

The LCLC Algorithm

To overcome these limitations, this paper introduces Learning from Common Local Clusters (LCLC). Unlike traditional instance-based methods that expand the positive set point-by-point, LCLC operates on a cluster level, making it significantly more robust to noise and outliers.

The algorithm begins by initializing a "pure" positive set ($P$) from a minimal seed. It then partitions unlabeled data into local clusters and applies Singular Value Decomposition (SVD) to extract Descriptive Common Principal Components (DCPCs). This feature selection step is vital for handling the high redundancy typical of time series data.

The LCLC cluster chain approach for decision boundary identification.

The core novelty of LCLC lies in its "cluster chain" mechanism. By constructing paths from the positive set ($P$) through ambiguous clusters ($AMBI$) to reliable negatives ($RN$), the algorithm identifies the "weakest link"—the maximum spatial gap between adjacent clusters. This gap serves as the natural decision boundary, allowing for a clean separation of likely positives and negatives.

Experimental Evaluation

LCLC was evaluated on five diverse datasets from the UCR archive: ECG (medical), Word Spotting (multimedia), Wafer (manufacturing), Yoga (image kinematics), and CBF (simulated data). To simulate extreme data scarcity, training was initiated with only a single positive seed instance, with all other training data treated as unlabeled.

Results and Discussion

Experimental results demonstrate that LCLC consistently outperforms state-of-the-art PU methods. As shown in the comparison below, LCLC achieved substantial gains in F-measure across all tested domains.

Performance comparison of LCLC vs. baseline methods.

Quantitatively, LCLC reached F-measures as high as 86.7% for ECG and 85.4% for Yoga. Compared to the second-best baseline (Ratana’s method), LCLC provided absolute improvements ranging from 2.7% to over 39%. Ablation studies further confirmed that both the DCPC-based feature selection and the cluster chaining mechanism are indispensable; removing either component led to performance drops of up to 40.4%.

Conclusion

LCLC represents a significant advancement in PU learning for time series. By shifting from instance-based to cluster-based extraction and leveraging topological "chains" to find decision boundaries, it achieves high recall without sacrificing precision. Future work will explore probabilistic representations for ambiguous clusters to better handle boundary-zone uncertainty.

Summary updated

BibTeX
@inproceedings{nguyen2011positive,
  author = {Minh Nhut Nguyen and Xiaoli Li and See-Kiong Ng},
  title = {{Positive Unlabeled Leaning for Time Series Classification}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2011},
  doi = {10.5591/978-1-57735-516-8/ijcai11-240},
}
Figure