Seminar: Self-Tuning One-Class Support Vector Machines for Data Classification

Yiming Qian
M.Sc. Candidate
Supervisor: Dr. Minglun Gong

Self-Tuning One-Class Support Vector Machines for Data Classification

Department of Computer Science
Thursday, April 10, 2014, 11:30 a.m., Room EN 2022


Support Vector Machine (SVM) based classifiers are most popular approaches for data classification in machine learning. To obtain high classification accuracy, parameter tuning methods such as cross-validation are often applied. However, parameter tuning is a very time consuming process. To address this problem, a simple, efficient and parameter-free approach is proposed for both binary and multiclass classification problems in this thesis, and is especially useful when dealing with large-scale datasets in the presence of label noise. Grown out of one-class SVM, our approach enjoys several distinct features: First, by utilizing the advantage of online learning, it works well with dynamic data; Second, its decision boundary is learned based on both positive and negative examples, whereas the original one-class SVM training is only based on positive examples; Third, the internal parameters and especially the kernel bandwidth in the training process are self-tuned, which makes our approach handy to use even for first-time users.

The proposed approach is compared side-by-side with LIBSVM, arguably the most widely-used data classification system, in a sequence of empirical evaluations, where our approach is shown to perform almost as well as their optimal parameter settings tuned for individual datasets, while consuming only a fraction of the processing time. In addition, real-world problems in computer vision - namely foreground segmentation and boundary matting for live videos - are solved to demonstrate the ability on working on dynamic and large-scale datasets. Our approach is shown to be particularly competent at processing a wide range of videos under complex scenes, and near real-time processing speed (14 frames per second (FPS) without matting and 8 FPS with matting on a mid-range PC & GPU) is achieved for VGA-sized videos.

*Key Words:* Data Classification, Support Vector Machine, One-Class SVM, Parameter Free, Label Noise, Foreground Segmentation, Video Matting