Mining concept-drifting data streams using ensemble classifiers

Wang, Haixun; Fan, Wei; Yu, Philip S.; Han, Jiawei

doi:10.1145/956755.956778

Cited by 473 publications

(457 citation statements)

References 0 publications

Supporting

Mentioning

437

Contrasting

Unclassified

Order By: Relevance

“…As a result, the discriminative power of a hypothesis should be determined by its performance on each instance in

{scriptL}_{h o}^{(t)}

, rather than its overall performance on the entire set of

{scriptL}_{h o}^{(t)}

. This design takes the class imbalance issue into account and is different from other practices dealing with balanced data streams [14]. In the implementation, we use the logistic loss function, i.e., Δ( f ( x ), y ) = log(1 + exp(− yf ( x ))).…”

Section: Methodsmentioning

confidence: 99%

Classifying Imbalanced Data Streams via Dynamic Feature Group Weighting with Importance Sampling

Wu¹,

Edwards²,

Wang³

et al. 2014

Proceedings of the 2014 SIAM International Conference on Data Mining

Self Cite

View full text Add to dashboard Cite

Data stream classification and imbalanced data learning are two important areas of data mining research. Each has been well studied to date with many interesting algorithms developed. However, only a few approaches reported in literature address the intersection of these two fields due to their complex interplay. In this work, we proposed an importance sampling driven, dynamic feature group weighting framework (DFGW-IS) for classifying data streams of imbalanced distribution. Two components are tightly incorporated into the proposed approach to address the intrinsic characteristics of concept-drifting, imbalanced streaming data. Specifically, the ever-evolving concepts are tackled by a weighted ensemble trained on a set of feature groups with each sub-classifier (i.e. a single classifier or an ensemble) weighed by its discriminative power and stable level. The un-even class distribution, on the other hand, is typically battled by the sub-classifier built in a specific feature group with the underlying distribution rebalanced by the importance sampling technique. We derived the theoretical upper bound for the generalization error of the proposed algorithm. We also studied the empirical performance of our method on a set of benchmark synthetic and real world data, and significant improvement has been achieved over the competing algorithms in terms of standard evaluation metrics and parallel running time. Algorithm implementations and datasets are available upon request.

show abstract

“…As a result, the discriminative power of a hypothesis should be determined by its performance on each instance in

{scriptL}_{h o}^{(t)}

, rather than its overall performance on the entire set of

{scriptL}_{h o}^{(t)}

Section: Methodsmentioning

confidence: 99%

Classifying Imbalanced Data Streams via Dynamic Feature Group Weighting with Importance Sampling

Wu¹,

Edwards²,

Wang³

et al. 2014

Proceedings of the 2014 SIAM International Conference on Data Mining

Self Cite

View full text Add to dashboard Cite

show abstract

“…To overcome this problem, Street proposed a method to retiring old classifiers one at a time [37]. Wang assigned weights to classifiers proportional to their accuracy on the most recent data block [38]. Chu viewed the choice of weights as an optimization problem and used logistic regression to settle it [39].…”

Section: Challenges Of Traffic Identificationmentioning

confidence: 99%

Universal Feature Extraction for Traffic Identification of the Target Category

et al. 2016

View full text Add to dashboard Cite

Traffic identification of the target category is currently a significant challenge for network monitoring and management. To identify the target category with pertinence, a feature extraction algorithm based on the subset with highest proportion is presented in this paper. The method is proposed to be applied to the identification of any category that is assigned as the target one, but not restricted to certain specific category. We divide the process of feature extraction into two stages. In the stage of primary feature extraction, the feature subset is extracted from the dataset which has the highest proportion of the target category. In the stage of secondary feature extraction, the features that can distinguish the target and interfering categories are added to the feature subset. Our theoretical analysis and experimental observations reveal that the proposed algorithm is able to extract fewer features with greater identification ability of the target category. Moreover, the universality of the proposed algorithm proves to be available with the experiment that every category is set to be the target one.

show abstract

“…Vergara et al used a static ensemble of multiple SVMs to cope with the problem of drift in chemical gas sensors [26]. Wang et al also proposed a static ensemble of SVMs, which is similar to that of Vergara et al , only instead of the weight assignment method [27]. Amini et al used an ensemble of SVMs or MLPs on data from a single metal oxide gas sensor (SP3-AQ2, FIS Inc., Hyogo, Japan) operated at six different rectangular heating voltage pulses (temperature modulation), to identify analytes regardless of concentration [28].…”

Section: Related Workmentioning

confidence: 99%

“…Wang et al [27] use the weight, β i = MSE r – MSE i , for each classifier, f i , where MSE i is the mean square error of classifier f i on S T and MSE r is the mean square error of a classifier predicting randomly.…”

Section: Dynamic Classifier Ensemble and Predictionsmentioning

confidence: 99%

“…Most classifier ensemble methods, such as those described in references [26,27], are proposed based on an assumption that the distribution of the examples in the following batch, S T +1 , does not change significantly from that in the current batch, S T . Thus, these methods use the examples in batch S T to estimate the weights, { β 1 , …, β T }, for S T +1 .…”

Section: Dynamic Classifier Ensemble and Predictionsmentioning

confidence: 99%

See 1 more Smart Citation

Metal Oxide Gas Sensor Drift Compensation Using a Dynamic Classifier Ensemble Based on Fitting

Hang

Tang

2013

Sensors

View full text Add to dashboard Cite

Sensor drift is currently the most challenging problem in gas sensing. We propose a novel ensemble method with dynamic weights based on fitting (DWF) to solve the gas discrimination problem, regardless of the gas concentration, with high accuracy over extended periods of time. The DWF method uses a dynamic weighted combination of support vector machine (SVM) classifiers trained by the datasets that are collected at different time periods. In the testing of future datasets, the classifier weights are predicted by fitting functions, which are obtained by the proper fitting of the optimal weights during training. We compare the performance of the DWF method with that of competing methods in an experiment based on a public dataset that was compiled over a period of three years. The experimental results demonstrate that the DWF method outperforms the other methods considered. Furthermore, the DWF method can be further optimized by applying a fitting function that more closely matches the variation of the optimal weight over time.

show abstract

Mining concept-drifting data streams using ensemble classifiers

Cited by 473 publications

References 0 publications

Classifying Imbalanced Data Streams via Dynamic Feature Group Weighting with Importance Sampling

Classifying Imbalanced Data Streams via Dynamic Feature Group Weighting with Importance Sampling

Universal Feature Extraction for Traffic Identification of the Target Category

Metal Oxide Gas Sensor Drift Compensation Using a Dynamic Classifier Ensemble Based on Fitting

Contact Info

Product

Resources

About