Title: Active learning with partially featured data
Abstract: In this paper, we propose a new active learning algorithm in which the learner chooses the samples to be queried from the unlabeled data points whose attributes are only partially observed. In addition, we propose a cost-driven decision framework where the learner chooses to query either the labels or the missing attributes. This problem statement addresses a common constraint when building large datasets and applying active learning techniques on them, where some of the attributes (including the labels) are significantly harder or more costly to acquire per data point. We take a novel approach to this problem, first by building an imputation model that maps from the partially featured data to the fully featured dimension, and then performing active learning on the projected input space combined with the estimated confidence of inference. We discuss that our approach is flexible and can work with graph mining tasks as well as conventional semi-supervised learning problems. The results suggest that the proposed algorithm facilitates more cost-efficient annotation than the baselines.
Publication Year: 2014
Publication Date: 2014-04-07
Language: en
Type: article
Indexed In: ['crossref']
Access and Citation
Cited By Count: 10
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot