FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W310295632> ?p ?o ?g. }

• W310295632 abstract "Instance ranking is a subfield of supervised machine learning and is concerned with inferring predictive models that can rank a set of data instances. We focus on multipartite ranking, where instances belong to one of a limited set of rank classes, study different approaches on synthetic and real data sets, and propose a ranking-specific evaluation framework and a new learning approach that combines multitask learning and binary decomposition.The thesis starts with an analysis of existing ranking approaches. These are used in a practical application of ranking within the domain of molecular biology. In particular, we study embryonic stem cell differentiation posed as a multipartite ranking problem. We critically evaluate several ranking approaches and demonstrate how they can be used to construct accurate predictive models that can rank samples based on their stage of differentiation. For testing, we introduce a framework for evaluation of ranking methods including a generalization of the popular performance measure AUC (area under the ROC curve) that can be used for multipartite problems. We proceed by analysing the family of methods based on binary decomposition, which reduces a ranking problem to a set of binary classification tasks. To improve the process of learning models for these tasks, we suggest to combine it with a multitask learning technique. Specifically, we propose a new ranking method, which we name BDMT, that combines one-against-one decomposition and multitask feature learning. The decomposition allows us to model nonlinear patterns and to simplify the learning domain to problems that are suitable for classical machine learning approaches. Multitask learning allows us to generalize across the decomposed tasks, making them interdependent through a joint regularized optimization. Our experiments show that the addition of multitask learning can greatly improve the performance of one-against-one decomposition and even succeed in outperforming state-of-the-art ranking approaches in certain settings. Learning the models of decomposed tasks simultaneously, increases the stability of model estimation and reduces the sensitivity to perturbations of the training data set. Compared with other ranking methods that are also able to model complex patterns, BDMT remains efficient and can achieve low training times with the use of fast linear base learners. We also show how the method and the patterns it learns can be interpreted. New features learned during the training of BDMT can reveal important hidden factors and hence map the problem into a low-dimensional subspace spanned by a set of novel features. Individual models of decomposed tasks and the similarities between them can be studied to further elucidate the distribution of rank classes in this subspace." @default.
• W310295632 created "2016-06-24" @default.
• W310295632 creator A5033376376 @default.
• W310295632 date "2014-04-03" @default.
• W310295632 modified "2023-09-27" @default.
• W310295632 title "Ranking by Multitask Learning" @default.
• W310295632 hasPublicationYear "2014" @default.
• W310295632 type Work @default.
• W310295632 sameAs 310295632 @default.
• W310295632 citedByCount "0" @default.
• W310295632 crossrefType "dissertation" @default.
• W310295632 hasAuthorship W310295632A5033376376 @default.
• W310295632 hasConcept C114614502 @default.
• W310295632 hasConcept C119857082 @default.
• W310295632 hasConcept C12267149 @default.
• W310295632 hasConcept C124975894 @default.
• W310295632 hasConcept C134306372 @default.
• W310295632 hasConcept C154945302 @default.
• W310295632 hasConcept C162324750 @default.
• W310295632 hasConcept C164226766 @default.
• W310295632 hasConcept C177148314 @default.
• W310295632 hasConcept C177264268 @default.
• W310295632 hasConcept C187736073 @default.
• W310295632 hasConcept C189430467 @default.
• W310295632 hasConcept C199360897 @default.
• W310295632 hasConcept C2780451532 @default.
• W310295632 hasConcept C28006648 @default.
• W310295632 hasConcept C33923547 @default.
• W310295632 hasConcept C36503486 @default.
• W310295632 hasConcept C41008148 @default.
• W310295632 hasConcept C48372109 @default.
• W310295632 hasConcept C66905080 @default.
• W310295632 hasConcept C86037889 @default.
• W310295632 hasConcept C94375191 @default.
• W310295632 hasConceptScore W310295632C114614502 @default.
• W310295632 hasConceptScore W310295632C119857082 @default.
• W310295632 hasConceptScore W310295632C12267149 @default.
• W310295632 hasConceptScore W310295632C124975894 @default.
• W310295632 hasConceptScore W310295632C134306372 @default.
• W310295632 hasConceptScore W310295632C154945302 @default.
• W310295632 hasConceptScore W310295632C162324750 @default.
• W310295632 hasConceptScore W310295632C164226766 @default.
• W310295632 hasConceptScore W310295632C177148314 @default.
• W310295632 hasConceptScore W310295632C177264268 @default.
• W310295632 hasConceptScore W310295632C187736073 @default.
• W310295632 hasConceptScore W310295632C189430467 @default.
• W310295632 hasConceptScore W310295632C199360897 @default.
• W310295632 hasConceptScore W310295632C2780451532 @default.
• W310295632 hasConceptScore W310295632C28006648 @default.
• W310295632 hasConceptScore W310295632C33923547 @default.
• W310295632 hasConceptScore W310295632C36503486 @default.
• W310295632 hasConceptScore W310295632C41008148 @default.
• W310295632 hasConceptScore W310295632C48372109 @default.
• W310295632 hasConceptScore W310295632C66905080 @default.
• W310295632 hasConceptScore W310295632C86037889 @default.
• W310295632 hasConceptScore W310295632C94375191 @default.
• W310295632 hasLocation W3102956321 @default.
• W310295632 hasOpenAccess W310295632 @default.
• W310295632 hasPrimaryLocation W3102956321 @default.
• W310295632 hasRelatedWork W115818832 @default.
• W310295632 hasRelatedWork W1505119859 @default.
• W310295632 hasRelatedWork W1687165158 @default.
• W310295632 hasRelatedWork W1758923908 @default.
• W310295632 hasRelatedWork W1773321559 @default.
• W310295632 hasRelatedWork W1978830285 @default.
• W310295632 hasRelatedWork W2104261769 @default.
• W310295632 hasRelatedWork W2108862644 @default.
• W310295632 hasRelatedWork W2145474226 @default.
• W310295632 hasRelatedWork W2305870132 @default.
• W310295632 hasRelatedWork W245305909 @default.
• W310295632 hasRelatedWork W251250361 @default.
• W310295632 hasRelatedWork W2744295891 @default.
• W310295632 hasRelatedWork W2774808155 @default.
• W310295632 hasRelatedWork W2965435631 @default.
• W310295632 hasRelatedWork W2984692558 @default.
• W310295632 hasRelatedWork W3100625157 @default.
• W310295632 hasRelatedWork W3141015273 @default.
• W310295632 hasRelatedWork W72335511 @default.
• W310295632 hasRelatedWork W2571941834 @default.
• W310295632 isParatext "false" @default.
• W310295632 isRetracted "false" @default.
• W310295632 magId "310295632" @default.
• W310295632 workType "dissertation" @default.