Paper Info
Title
Learning Feature Weights from Positive Cases.
Abstract
The availability of new data sources presents both opportunities and challenges for the use of Case-based Reasoning to solve novel problems. In this paper, we describe the research challenges we faced when trying to reuse experiences of successful academic collaborations available online in descriptions of funded grant proposals. The goal is to recommend the characteristics of two collaborators to complement an academic seeking a multidisciplinary team; the three form a collaboration that resembles a configuration that has been successful in securing funding. While seeking a suitable measure for computing similarity between cases, we were confronted with two challenges: a problem context with insufficient domain knowledge and data that consists exclusively of successful collaborations, that is, it contains only positive instances. We present our strategy to overcome these challenges, which is a clustering-based approach to learn feature weights. Our approach identifies poorly aligned cases, i.e., ones that violate the assumption that similar problems have similar solutions. We use the poorly aligned cases as negatives in a feedback algorithm to learn feature weights. The result of this work is an integration of methods that makes CBR useful to yet another context and in conditions it has not been used before.
Year
DOI
Venue
2013
10.1007/978-3-642-39056-2_10
ICCBR
Keywords
Field
DocType
Case Alignment,Case Cohesion,Density Clustering,Multidisciplinary Collaboration,Recommender Systems,Single Class Learning,Subspace Clustering
Recommender system,Data mining,Problem context,Yet another,Domain knowledge,Multidisciplinary approach,Reuse,Computer science,Knowledge management,Cluster analysis,Multidisciplinary Collaboration
Conference
Citations 
PageRank 
References 
1
0.35
19
Authors
3
Name
Order
Citations
PageRank
Sidath Gunawardena182.16
Rosina Weber233434.42
Stoyanovich Julia334136.67