Discriminant Independent Vector Analysis

Abstract

With the rapid advancement of technology and the accelerated growth in data production, Blind Source Separation (BSS) methods have gained increasing relevance due to their broad applicability across diverse domains. In scenarios where multiple data modalities are mixed and the objective is to recover the original underlying sources, techniques capable of exploiting relationships among them become essential, particularly in representation and classification tasks. The traditional method for handling such multimodal data is Independent Vector Analysis (IVA); however, its non-discriminative nature limits its performance when source separation is directly linked to classification objectives. In this context, this work introduces Discriminant Independent Vector Analysis (DIVA), a supervised extension of IVA constructed through the incorporation of Fisher’s Linear Discriminant (FLD) criterion into the IVA framework. The resulting method aims to estimate independent sources that, in addition to satisfying statistical independence, maximize class separability, making it particularly suitable for binary classification problems. The proposed model was implemented based on IVA-G, a widely established formulation in the literature. To evaluate the performance of DIVA-G, a Support Vector Machine (SVM) classifier was employed in a semi-supervised setting, using the F1-score as the primary evaluation metric. Experiments with synthetic datasets enabled the identification of statistical characteristics that favor the method, demonstrating consistent and superior performance compared with other IVA-derived algorithms. Subsequently, the real-world datasets NSL-KDD and MediaEval2016 were used to investigate the behavior of the method in complex and noisy scenarios. The results indicate satisfactory performance, stability, and reliability, suggesting that DIVA-G is a promising approach for discriminative source separation and warrants further, more in-depth investigation.