Abstract:Linear discriminant analysis (LDA) represents a simple yet powerful technique for partitioning a p-dimensional feature vector into one of K classes based on a linear projection learned from N labeled observations. However, it is well-established that in the high-dimensional setting (p > N) the underlying projection estimator degenerates. Moreover, any linear discriminate function involving a large number of features may be difficult to interpret. To ameliorate these issues, two general categories of sparse LDA modifications have been proposed, both to reduce the number of active features and to stabilize the resulting projections. The first, based on optimal scoring, is more straightforward to implement and analyze but has been heavily criticized for its ambiguous connection with the original LDA formulation. In contrast, a second strategy applies sparse penalty functions directly to the original LDA objective but requires additional heuristic trade-off parameters, has unknown global and local minima properties, and requires a greedy sequential optimization procedure. In all cases the choice of sparse regularizer can be important, but no rigorous guidelines have been provided regarding which penalty might be preferable. Against this backdrop, we winnow down the broad space of candidate sparse LDA algorithms and promote a specific selection based on optimal scoring coupled with a particular, complementary sparse regularizer. This overall process ultimately progresses our understanding of sparse LDA in general, while leading to targeted modifications of existing algorithms that produce superior results in practice on three high-dimensional gene data sets.

A Study on Decoupled Probabilistic Linear Discriminant Analysis

Discriminative analysis of direct linear discriminant analysis for face recognition

Local Pairwise Linear Discriminant Analysis for Speaker Verification

Local Learning Multiple Probabilistic Linear Discriminant Analysis.

Joint Probabilistic Linear Discriminant Analysis

Single-sided approach to discriminative PLDA training for text-independent speaker verification without using expanded i-vector

Minimum divergence estimation of speaker prior in multi-session PLDA scoring

Neural Discriminant Analysis for Deep Speaker Embedding

Local training for PLDA in speaker verification

Adaptation of PLDA for Multi-Source Text-Independent Speaker Verification

I-vector Kullback-Leibler divisive normalization for PLDA speaker verification.

Channel Adaptation of Plda for Text-Independent Speaker Verification.

Multiobjective Optimization Training of PLDA for Speaker Verification

Deep Neural Network Based Discriminative Training for I-Vector/PLDA Speaker Verification.

Constrained Linear Discrimination Analysis and Face Recognition

Improving Short Utterance PLDA Speaker Verification using SUV Modelling and Utterance Partitioning Approach

Weakly Supervised PLDA Training.

Probabilistic Spherical Discriminant Analysis: An Alternative to PLDA for length-normalized embeddings

Understanding and Evaluating Sparse Linear Discriminant Analysis

Improved Supervised Locality Preserving Projection for I-vector Based Speaker Verification

Domain adaptation based Speaker Recognition on Short Utterances