Non-Overlapped Multi-View Weak-Label Learning Guided by Multiple Correlations
Abstract: Insufficient labeled training samples pose a critical challenge in multi-label classification, potentially leading to overfitting of the model. This paper delineates a criterion for establishing a common domain among different datasets, whereby datasets sharing analogous object descriptions and label structures are considered part of the same field. Integrating samples from disparate datasets within this shared field for training purposes effectively mitigates overfitting and enhances model accuracy. Motivated by this approach, we introduce a novel method for multi-label classification termed Non-Overlapped Multi-View Weak-Label Learning Guided by Multiple Correlations (NOMWM). Our method strategically amalgamates samples from diverse datasets within the shared field to enrich the training dataset. Furthermore, we project samples from various datasets onto a unified subspace to facilitate learning in a consistent latent space. Additionally, we address the challenge of weak labels stemming from incomplete label overlaps across datasets. Leveraging weak-label indicator matrices and label correlation mining techniques, we effectively mitigate the impact of weak labels. Extensive experimentation on multiple benchmark datasets validates the efficacy of our method, demonstrating clear improvements over existing state-of-the-art approaches.
Primary Subject Area: [Content] Multimodal Fusion
Secondary Subject Area: [Experience] Multimedia Applications
Relevance To Conference: In practical application scenarios, numerous samples within multimedia or multimodal datasets often exhibit polysemy, where a research subject is associated with multiple labels, which frequently exhibit certain correlations. For instance, a musical composition may encompass elements such as "symphony," "Beethoven," "piano," and "classical music." To address classification challenges akin to the aforementioned scenario, multi-label learning frameworks have emerged. The objective of multi-label learning is to train an effective model capable of assigning all relevant labels to an unknown sample. If datasets describe similar objects and share akin label sets, they are designated as belonging to the "same field." The method proposed in this paper enhances training data by amalgamating samples from "same field" multimodal datasets and maps them to a unified subspace for multi-label classification learning, demonstrating broad applicability in the multimedia or multimodal domain.
Submission Number: 894
Loading