Amitava Karmaker

A boosting approach to remove class label noise

Ensemble methods have been known to improve prediction accuracy over the base learning algorithm. AdaBoost is well-recognized for that in its class. However, it is susceptible to overfitting the training instances corrupted by class label noise. This paper proposes a modification to AdaBoost that is more tolerant to class label noise, which further enhances its ability to boost prediction accuracy. In particular, we observe that in Adaboost, the weight-hike of noisy examples can be constrained by careful application of a cut-off in their weights. Effectiveness of our algorithm is demonstrated empirically using some artificially generated data. We also corroborate this on a number of data sets from UCI repository (Blake and Mertz, 1998). In both experimental settings, the results obtained affirm the efficacy of our approach. Finally, some of the significant characteristics of our technique related to noisy environments have been investigated.

Incorporating an EM-approach for handling missing attribute-values in decision tree induction

Data with missing attribute-values are quite common in many classification problems. In this paper, we incorporate an expectation-maximization (EM) inspired approach for filling up missing values to decision tree learning with the objective of improving classification accuracy. Here, each missing attribute-value is iteratively filled using a predictor constructed from the known values and predicted values of the missing attribute-values from the previous iteration. We show that our approach significantly outperforms some standard machine learning methods for handling missing values in classification tasks.

iBoost: Boosting using an instance-based exponential weighting scheme

AdaBoost is a well-recognized ensemble method to improve prediction accuracy over the base learning algorithm. However, it is prone to overfitting the training instances [18]. Freund, Mansour and Schapire [5] established that using exponential weighting scheme in combining classifiers reduces the problem of overfitting. Also, Helmbold, Kwek and Pitt [7] showed in the prediction using a pool of experts framework an instance-based weighting scheme improves performance. Motivated by these results, we propose here an instance-based exponential weighting scheme in which the weights of the base classifiers are adjusted according to the test instance x. Here, a competency classifier c_i is constructed for each base classifier h_i to predict whether the base classifiers guess of xs label can be trusted and adjust the weight of h_i accordingly. We show that this instance-based exponential weighting scheme enhances the performance of AdaBoost.

Identifying Transcription Regulatory Elements in the Human and Mouse Genomes Using Tissue-specific Gene Expression Profiles

Summary Revealing the complex interaction between trans- and cis-regulatory elements and identifying these potential binding sites are fundamental problems in understanding gene expression. The progresses in ChIP-chip technology facilitate identifying DNA sequences that are recognized by a specific transcription factor. However, protein-DNA binding is a necessary, but not sufficient, condition for transcription regulation. We need to demonstrate that their gene expression levels are correlated to further confirm regulatory relationship. Here, instead of using a linear correlation coefficient, we used a non-linear function that seems to better capture possible regulatory relationships. By analyzing tissue-specific gene expression profiles of human and mouse, we delineate a list of pairs of transcription factor and gene with highly correlated expression levels, which may have regulatory relationships. Using two closely-related species (human and mouse), we perform comparative genome analysis to cross-validate the quality of our prediction. Our findings are confirmed by matching publicly available TFBS databases (like TRANFAC and ConSite) and by reviewing biological literature. For example, according to our analysis, 80% and 85.71% of the targets genes associated with E2F5 and RELB transcription factors have the corresponding known binding sites. We also substantiated our results on some oncogenes with the biomedical literature. Moreover, we performed further analysis on them and found that BCR and DEK may be regulated by some common transcription factors. Similar results for BTG1, FCGR2B and LCK genes were also reported.