Abstract
Detecting outliers in multi-view data is a complex task due to the presence of different types of anomalies. Traditional methods often focus on either class-outliers (inconsistencies across views) or attribute-outliers (abnormalities within views), but rarely both. This paper introduces DMOD (Dual-regularized Multi-view Outlier Detection), a novel framework that simultaneously identifies both types of anomalies by decomposing data into latent cluster assignments and sample-specific errors. Experimental results across multiple benchmark datasets demonstrate that DMOD significantly outperforms state-of-the-art methods, achieving substantial gains in detection accuracy.
Introduction
In many real-world applications, data is collected from multiple sources or "views." For instance, a person's identity can be represented by facial images and voice recordings. Outlier detection in such multi-view contexts is challenging because an object might appear normal in one view but reveal its anomalous nature when compared across views.
Existing research typically addresses two distinct categories of outliers:
- Class-outliers: Samples that exhibit inconsistent cluster memberships across different views (e.g., appearing as "Class A" in View 1 but "Class B" in View 2).
- Attribute-outliers: Samples that are inherently abnormal within individual views, such as corrupted data points.
Most current multi-view methods prioritize class-outliers, while single-view methods handle attribute-outliers. DMOD bridges this gap by providing a unified framework to detect both simultaneously.
Methodology: The DMOD Framework
The DMOD framework utilizes matrix factorization to represent data from each view $i$ as: $$X^{(i)} = H^{(i)}G^{(i)} + S^{(i)}$$
Detecting Class-Outliers
The term $G^{(i)}$ represents latent coefficients corresponding to cluster memberships. DMOD employs a dual-regularization term to align these memberships across views using an alignment matrix $M_{ij}$. This approach accounts for the "orderless" nature of clusters (where Cluster 1 in one view might correspond to Cluster 3 in another). Samples that cannot be consistently aligned across views are flagged as class-outliers.
Detecting Attribute-Outliers
The term $S^{(i)}$ captures sample-specific errors or "weirdness." By applying an $\ell_{2,1}$-norm to this error matrix, the model encourages row-sparsity, forcing most errors to zero. Non-zero rows effectively highlight attribute-outliers that deviate from the expected cluster structure in the original feature space.
Outlier Measurement Criterion
The authors propose a unified Outlier Measurement Criterion ($\phi$) that integrates inconsistency in latent cluster assignments and the magnitude of error terms into a single outlier score for each sample.
Experimental Evaluation
DMOD was evaluated on five benchmark datasets, including UCI datasets (iris, breast cancer, ionosphere, letter) and the BUAA VisNir Face Database. Outliers were manually injected to test the model's robustness: class-outliers were created by swapping feature subsets, and attribute-outliers were generated using random values.
Performance Results
DMOD consistently outperformed state-of-the-art baselines, including HOAD (multi-view) and LRR (single-view).
On the letter dataset, DMOD achieved a remarkable 26.13% improvement in Area Under the Curve (AUC) over the next best method. Unlike specialized methods that excel only in specific outlier scenarios, DMOD maintained dominant performance regardless of the mix of class and attribute outliers.
Discussion and Future Work
While DMOD is highly effective and scales linearly with the number of samples, it does require pre-setting the number of clusters ($K$). Sensitivity analysis shows that setting $K$ too low can lead to information loss and decreased performance. Future research could focus on automating the selection of $K$ and refining the thresholding process for binary outlier classification.
Conclusion
DMOD represents a significant advancement in multi-view outlier detection by treating class and attribute anomalies as distinct but related phenomena. By operating in both latent and original feature spaces, it provides a robust and comprehensive solution for complex real-world data analysis.
Summary updated