Conference Review Session


A New Boosting Algorithm Using Input-Dependent Regularizer
ICML-2003
Yan Liu

AdaBoost has proved to be an effective method to improve the performance of base classifiers both theoretically and empirically. However, previous studies have shown that AdaBoost might suffer from the overfitting problem, especially for noisy data. In addition, most current work on boosting assumes that the combination weights are fixed constants and therefore does not take particular input patterns into consideration. In this paper, we present a new boosting algorithm, "WeightBoost", which tries to solve these two problems by introducing an input-dependent regularization factor to the combination weight. Similarly to AdaBoost, we derive a learning procedure for WeightBoost, which is guaranteed to minimize training errors. Empirical studies on eight different UCI data sets and one text categorization data set show that WeightBoost almost always achieves a considerably better classification accuracy than AdaBoost. Furthermore, experiments on data with artificially controlled noise indicate that the WeightBoost is more robust to noise than AdaBoost.


Improving the Graph Mincut Approach to Learning from Labeled and Unlabeled Examples
ICML-2003
Mugizi Robert Rwebangira

We present work in progress on the problem of learning from labeled and unlabeled data. Following the approach of Blum and Chawla (2001) we consider algorithms that use pairwise relationships among examples (both labeled and unlabeled) in order to construct a graph that can then be partitioned using a mincut algorithm. In our work, we extend the Blum-Chawla approach by adding noise to the graph structure and using this noise to provide the algorithm with an estimate of its own confidence. This gives the mincut approach some of the properties of the label propagation method of Zhu and Ghahramani (2002). We present preliminary results of this approach.


Sample Complexity When Learning from Labeled and Unlabeled Data
ICML-2003
Maria Florina Balcan

This talk presents an overview of our research in developing a PAC style framework for learning with both labeled and unlabeled data. To do this we enrich the standard PAC model with a notion of compatibility between a concept and a distribution, for expressing our belief that the target concept is compatible with the distribution over the unlabeled data. In a sense, this is like modifying the notion of a Concept Class from a set of concepts to a set of (concept, distribution) pairs. We analyze for different settings how much unlabeled data and how much labeled data do we need in order to identify with high confidence a hypothesis that is very close to the target concept we are searching for. We are mainly considering using unlabeled data for adjusting our prior, more specifically for ordering the functions in the hypothesis space (with respect to their degree of compatibility with the distribution over the unlabeled data) and we are dealing only with situations where the compatibility can be estimated as an expectation over unlabeled examples. We also study how the need for labeled data depends on the accuracy of our belief (that the target concept is compatible with the distribution over the unlabeled data). This is joint work with Avrim Blum.


Back to the Main Page

Charles Rosenberg
Last modified: Wed Sep 17 18:21:12 EDT 2003