Learning to Rank with Partially Labeled Training Data

Abstract : Many real life applications involve the ranking of objects instead of their classification. For example, in Document Retrieval the goal is to rank documents from a collection based on their relevancy to a user’s query. Recently the supervised learning of ranking functions has attracted considerable attention from the Machine Learning community and most computational models proposed for ranking rely on this paradigm. Labeling large amounts of data may require expensive human resources, especially for ranking problems, and they are unrealistic in most applications. In the other hand, the semi-supervised learning paradigm which considers the possibility of learning from both the labeled and unlabeled examples has attracted the interest of the ML community in the field of classification since 1998. In this paper, we propose a semi-supervised learning algorithm for ranking. Existing semi supervised ranking algorithms are graph-based transductive techniques which from an observed training dataset, order a specific unlabeled data pool. Our motivation here is to develop a novel inductive approach which from a specific observed training data (labeled and unlabeled) produces a general ranking rule, which ranks unseen examples with high accurancy. Our algorithm is an iterative approach which combines a supervised and a graph-based method. Empirical results on a real-life dataset from the CACM collection have shown the potential of this approach in the context of Document Retrieval.
Document type :
Conference papers
Complete list of metadatas

https://hal.archives-ouvertes.fr/hal-01337596
Contributor : Lip6 Publications <>
Submitted on : Monday, June 27, 2016 - 11:56:55 AM
Last modification on : Thursday, March 21, 2019 - 2:16:34 PM

Identifiers

  • HAL Id : hal-01337596, version 1

Citation

Tuong Vinh Truong, Massih-Reza Amini, Patrick Gallinari. Learning to Rank with Partially Labeled Training Data. 1st International Conference on Multidisciplinary Information Sciences and Technologies, Oct 2006, Merida, Spain. pp.64-74. ⟨hal-01337596⟩

Share

Metrics

Record views

147