Transcript CIT_b

Learning from Positive and Unlabeled Examples
Investigator: Bing Liu, Computer Science
Prime Grant Support: National Science Foundation
Problem Statement and Motivation
Positive
training data
Unlabeled
data
• Given a set of positive examples P and a set of unlabeled
examples U, we want to build a classifier.
• The key feature of this problem is that we do not have
labeled negative examples. This makes traditional
classification learning algorithms not directly applicable.
Learning algorithm
•.The main motivation for studying this learning model is to
solve many practical problems where it is needed. Labeling
of negative examples can be very time consuming.
Classifier
Technical Approach
We have proposed three approaches.
• Two-step approach: The first step finds some reliable
negative data from U. The second step uses an iterative
algorithm based on naïve Bayesian classification and
support vector machines (SVM) to build the final classifier.
• Biased SVM: This method models the problem with a
biased SVM formulation and solves it directly. A new
evaluation method is also given, which allows us to tune
biased SVM parameters.
• Weighted logistic regression: The problem can be
regarded as an one-side error problem and thus a weighted
logistic regress method is proposed.
Key Achievements and Future Goals
• In (Liu et al. ICML-2002), it was shown theoretically that
P and U provide sufficient information for learning, and
the problem can be posed as a constrained optimization
problem.
• Some of our algorithms are reported in (Liu et al. ICML2002; Liu et al. ICDM-2003; Lee and Liu ICML-2003; Li
and Liu IJCAI-2003).
• Our future work will focus on two aspects:
• Deal with the problem when P is very small
• Apply it to the bio-informatics domain. There are
many problems there requiring this type of learning.
Gene Expression Programming for Data Mining and
Knowledge Discovery
Investigators: Peter Nelson, CS; Xin Li, CS; Chi Zhou, Motorola Inc.
Prime Grant Support: Physical Realization Research Center of Motorola Labs
Problem Statement and Motivation
Genotype:
sqrt.*.+.*.a.*.sqrt.a.b.c./.1.-.c.d
Phenotype:
Mathematical form:
(a  bc)  a
1
cd
Figure 1. Representations of solutions in GEP
Technical Approach
• Overview: improving the problem solving ability of the
GEP algorithm by preserving and utilizing the selfemergence of structures during its evolutionary process
• Constant Creation Methods for GEP: local optimization
of constant coefficients given the evolved solution
structures to speed up the learning process.
• A new hierarchical genotype representation: natural
hierarchy in forming the solution and more protective
genetic operation for functional components
• Dynamic substructure library: defining and reusing selfemergent substructures in the evolutionary process.
• Real world data mining tasks: large data set, high
dimensional feature set, non-linear form of hidden
knowledge; in need of effective algorithms.
• Gene Expression Programming (GEP): a new
evolutionary computation technique for the creation of
computer programs; capable of producing solutions of
any possible form.
• Research goal: applying and enhancing GEP
algorithm to fulfill complex data mining tasks.
Key Achievements and Future Goals
• Have finished the initial implementation of the
proposed approaches.
• Preliminary testing has demonstrated the feasibility and
effectiveness of the implemented methods: constant
creation methods have achieved significant improvement
in the fitness of the best solutions; dynamic substructure
library helps identify meaningful building blocks to
incrementally form the final solution following a faster
fitness convergence curve.
• Future work include investigation for parametric
constants, exploration of higher level emergent
structures, and comprehensive benchmark studies.