Ent?te

Logo du LIFL

Depuis le 1er janvier 2015 le LIFL et le LAGIS forment le laboratoire CRIStAL

  1. Actualités

Séminaire de

Mark Zhang

21 juin 2010
Amphithéâtre de l'IRCICA

Multiple-Instance Learning on Multimodal Data Mining in a Multimedia Database

The multimodal data mining problems in a multimedia database, such as image annotation and
image retrieval, may be tackled as multiple instance learning problems. The multiple instance
learning problem can be converted to a standard supervised learning problem by the existing
methods. Following this line of research we propose a learning method, EMIL (Enhanced Multi-
ple Instance Learning), which combines the 1-norm SVM and the standard 2-norm SVM together.

The 1-norm SVM is applied to select important features, followed by the 2-norm SVM which works
only on the selected features. Furthermore, we present a semi-supervised learning technique called
Multiple-Instance Semi-Supervised Learning by Embedded Instance Selection (MIS3), which ex-
tends EMIL to incorporate the unlabeled data into the learning process. Unlike the existing semi-
supervised learning methods, MIS3 makes use of the unlabeled data in two steps. The instance-
based feature mapping which utilizes the unlabeled data converts an MIL problem to a supervised
learning problem, followed by a feature selection procedure. An existing semi-supervised learning
method is then adopted to utilize the unlabeled data again. MIS3 which naturally extends to
the out-of-sample data is an inductive learning method, excelling the existing multiple-instance
semi-supervised learning methods in the literature which are transductive in nature. We have
performed experiments on several multimedia databases for different data mining tasks and the
experimental results show significant performance improvements over state-of-the-art methods.
 

Ours

UMR 8022 - Laboratoire d'Informatique Fondamentale de Lille - Copyright © 2012 Sophie TISON - Crédits & Mentions légales

Page respectant XHTML et CSS.

Pour tout commentaire / Comments and remarks : webmaster