A showcase of ÉTS researchers’ publications and other contributions

Score thresholding for accurate instance classification in multiple instance learning


Downloads per month over past year

Carbonneau, Marc-André, Granger, Eric et Gagnon, Ghyslain. 2016. « Score thresholding for accurate instance classification in multiple instance learning ». In Sixth International Conference on Image Processing Theory, Tools and Applications (IPTA) (Oulu, Finland, Dec. 12-15, 2016) Piscataway, NJ, USA : IEEE.

[thumbnail of Score-Thresholding-for-Accurate-Instance-Classification-in-Multiple-Instance-Learning.pdf]

Download (723kB) | Preview


Multiple instance learning (MIL) is a form of weakly supervised learning for problems in which training instances are arranged into bags, and a label is provided for whole bags but not for individual instances. Most proposed MIL algorithms focused on bag classification, but more recently, the classification of individual instance has attracted the attention of the pattern recognition community. While these two tasks are similar, there are important differences in the consequences of instance misclassification. In this paper, the scoring function learned by MIL classifier for the bag classification task is exploited for instance classification by adjusting the decision threshold. A new criterion for the threshold adjustment is proposed and validated using 7 reference MIL algorithms on 3 realworld data sets from different application domains. Experiments show a considerable improvement in accuracy of these algorithms for instance classification. In some application, the unweighted average recall increases by as much as 18%, while the F1-score increases by 12%.

Item Type: Conference proceeding
ISBN: 978-1-4673-8910-5
Granger, Éric
Gagnon, Ghyslain
Affiliation: Génie électrique, Génie de la production automatisée
Date Deposited: 09 Feb 2017 21:51
Last Modified: 12 Apr 2017 20:42

Actions (login required)

View Item View Item