Anchor PCA

Researchers have proposed a new technique called Anchor PCA for robust unsupervised dimension reduction from multi-domain data, improving upon traditional Principal Component Analysis (PCA)^[1].

PCA is widely used for dimension reduction in machine learning-based systems, including cyberattack detection^[1]. However, when applied to data from multiple related domains, traditional PCA can focus on spurious directions that exhibit high variation in only a few domains. Anchor PCA addresses this issue by trading off overall explained variance with agreement between shared and domain-specific low-rank embeddings. This approach amounts to PCA on a modified target matrix and can be solved efficiently. According to the researchers, Anchor PCA recovers a maximal invariant subspace and yields embeddings that explain more variance on unseen domains than the pooling baseline and a worst-case alternative^[1]. Studies on simulated and real-world gas sensor data with temporal drift demonstrate the effectiveness of Anchor PCA. High-dimensional feature representations are common in machine learning-based cyberattack detection systems, and techniques like PCA can substantially reduce feature dimensionality with minimal impact on classification accuracy^[2].