.png){kind=icon}
.png){kind=icon}
Return
Scalable multi-label annotation via semi-supervised kernel semantic embedding
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patrec.2018.10.001.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In this paper, we present a novel semi-supervised semantic embedding method based on kernel matrix factorization for automatic multi-label annotation. The method called Semi-supervised Online Kernel Semantic Embedding (SS-OKSE) builds a semantic representation modeled by the document features and the associated labels when available, from this final semantic representation another transformation is simultaneously learned to predict the labels for new documents. Thanks to its kernel-based formulation, the proposed method is suitable for modeling non-linear complex relationships among the data samples. A scalable architecture based on a learning-on-a-budget strategy and its formulation as an end-to-end architecture allows the efficient training using online learning based on stochastic gradient descent. Another distinctive feature of the proposed method is that the kernel parameters can be learned directly, avoiding the necessity of an exhaustive exploration to determine their appropriate values. The effectiveness of the SS-OKSE method was evaluated in a multi-label annotation task under a semi-supervised learning setup over two different textual datasets and it was compared against several supervised and semi-supervised techniques. Experimental results show that SS-OKSE exhibits a significant improvement, showing that a better modeling can be achieved with an adequate selection/construction of a kernel input representation. (C) 2018 Published by Elsevier B.V.
Keywords:
Semantic representation
Semi-supervised learning
Learning on a budget
Kernel matrix factorization
Multi-label annotation
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
3.3
Papers:
7.8K
Citations:
1.6W