.png){kind=icon}
.png){kind=icon}
Return
Triple contrastive learning representation boosting for supervised multiclass tasks
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2024.104011.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Supervised contrastive learning is highly effective for extracting sample representations, thereby enhancing the performance of downstream tasks. However, existing methods underutilize supervised signals in multi-class datasets, leading to reduced inter-class distances and consequently weakening the model's generalization ability. To address this, we introduced a triple-supervised contrastive learning approach that expands the traditional loss function constraints of (anchor- positive, anchor-negative) to include a new (between-negative) constraint. This modification increases the separation between negatively labeled samples originating from the same anchor, enhancing model performance and generalization. We applied this approach to the three most commonly used supervised contrastive learning methods and conducted experiments across six text classification and two image classification datasets. Our evaluation involved measuring classification accuracy and macro-F1 scores, analyzing inter- and intra-class distances, and visualizing representation clusters. The introduction of our triple constraints resulted in average classification accuracy improvements of 0.43%, 0.35%, and 0.25% compared to the original methods, demonstrating the enhanced ability of our approach to leverage supervised signals effectively in multi-class datasets. Additionally, we incorporated our triple constraint into four state-of-the-art supervised contrastive learning methods from different domains, representing the latest advancements, to further demonstrate the wide applicability and effectiveness of our approach. Our baseline code, improved code, and datasets are all open-sourced as follows: https://github.com/6akso/Triple-Contrastive-Learning.git
Keywords:
Supervised contrastive learning
Multi-class datasets
Text classification
Image classification
Journal
I
IF:
6.9
Papers:
5.2K
Citations:
1.4W
Organization
No organization information available