Incremental learning for transductive support vector machine

doi:10.1016/j.patcog.2022.108982

2023-01-01

PRE

DOI：10.1016/j.patcog.2022.108982

原文链接

原文求助

摘要

En 中文

Semi-supervised learning is ubiquitous in real-world machine learning applications due to its good per-formance for handling the data where only a few number of samples are labeled while most of then are unlabeled. Transductive support vector machine (TSVM) is an important semi-supervised learning method which formulates the problem as a nonconvex combinatorial optimization problem. The infinitesimal an-nealing algorithm is a novel training method of TSVM which can alleviate the impact of the combinatorial and non-convex natures in TSVM and achieve a fast training of TSVM. However, it is still a challenging problem to handle large-scale data for TSVM even using the infinitesimal annealing algorithm. To miti-gate this problem, in this paper, we propose an incremental learning algorithm for TSVM (ILTSVM) based on the path following technique under the framework of infinitesimal annealing. Specifically, for new samples, we call CP-Step to change the solution and partition by increasing the size of the penalty co-efficient. The difference between training labeled samples and training unlabeled samples is that the variation range of the penalty coefficient of labeled samples is larger than that of unlabeled samples. If in the process of CP-Step, pseudo-labels of unlabeled samples are classified incorrectly, call DJ-Step to flip the pseudo-labels, and use incremental and decremental algorithms to make the KKT condition satisfied. We also analyze the time complexity and convergence of ILTSVM. The experimental results show that compared with other incremental or batch learning algorithms, our algorithm is the most effective and fastest method for training TSVM.(c) 2022 Published by Elsevier Ltd.

Keyword:

Transductive support vector machine

Incremental learning

Non -convex optimization

Infinitesimal annealing