Abstract
Graphs have a crucial impact on the performance of any graph-based semi-supervised learning method, so their construction should be carefully considered. In this letter, and in the context of semi-supervised learning, we will address graph-based semi-supervised learning employing multiple views for the data. The notion of data smoothness is another missing concept in graph construction that should be considered when constructing graphs. Compared to a single feature, using multiple sources of information can increase the efficiency of the post-processing task that adopts the constructed graph. Therefore, we present an approach that merges the notions of data smoothness and label smoothness with label fitness and projection matrix calculation. Moreover, two or more views are merged to exploit the information hidden in different features. Experiments performed with image databases show the superiority of the proposed approach compared to single features and other competing fusion algorithms. Compared to recent fusion methods, the introduced scheme improved the semi-supervised classification performance. For example, on the MNIST dataset with 20 labeled images per class, the average improvement due to the proposed labeling inference was 4.4%. The proposed method is inductive and computes a linear mapping to obtain the label of unseen or test patterns.
Fadi Dornaika
Ikerbasque Research Professor
Ikerbasque Research Professor with expertise in computer vision, machine learning, and pattern recognition.
Zoulfikar Ibrahim
Professor and Software Engineer
I develop scalable graph-based machine learning methods and teach cutting-edge technologies in web and software development. My focus spans semi-supervised learning, data analysis, and full-stack engineering.
Nagore Barrena
Assistant Professor
My research focuses on computer vision and deep learning, with applications to biomedical and natural images.