Tensor analysis with n-mode generalized difference subspace

The increasing use of multiple sensors, which produce a large amount multi-dimensional data, requires efficient representation and classification methods. In this paper, we present new method for data that relies on two premises: (1) are usually represented by tensors, since brings benefits from multilinear algebra established tensor factorization methods; (2) can be described subspace vector space. has been employed pattern-set recognition, its counterpart is also available in the literature. However, traditional methods do not discriminative information degrading accuracy. case, generalized difference (GDS) provides an enhanced reducing redundancy revealing structures. Since GDS does handle propose projection called n-mode GDS, efficiently handles data. We introduce Fisher score as class separability index improved metric based geodesic distance similarity. experimental results gesture action recognition show proposed outperforms commonly used literature without relying pre-trained models or transfer learning.