Chan, Philip and Salvatore Stolfo. 1997. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8(1):5--28.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... including bagging, arcing, boosting and correlation control [6, 31] Approaches to pooling classifiers can be separated into two main categories: i) simple combiners, e.g. voting [3] Bayesian based weighted product rule [22] or averaging [24, 30] and, ii) meta learners, such as arbitration [7] or stacking [34] The simple combining methods are best suited for problems where the individual classifiers perform the same task, and have comparable success. However, such combiners are more susceptible to outliers and to unevenly performing classifiers. In the second category, either sets of ....

P. Chan and S. Stolfo. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Integration of Information, 8(1):8--25, 1997.

....model assumes that all the data required by any data mining algorithm is either available at or can be sent to a central site. A simple approach to data mining over multiple sources that will not share data is to run existing data mining tools at each site independently and combine the results[5, 6, 17]. However, this will often fail to give globally valid results. Issues that cause a disparity between local and global results include: Values for a single entity may be split across sources. Data mining at individual sites will be unable to detect cross site correlations. The same item ....

....al. proposed a method for horizontally partitioned data[8] and more recent work has addressed privacy in this model[14] Distributed classification has also been addressed. A meta learning approach has been developed that uses classifiers trained at di#erent sites to develop a global classifier [6, 17]. This could protect the individual entities, but it remains to be shown that the individual classifiers do not disclose private information. Recent work has addressed classification using Bayesian Networks in vertically partitioned data [7] and situations where the distribution is itself ....

P. Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

....model assumes that all the data required by any data mining algorithm is either available at or can be sent to a central site. A simple approach to data mining over multiple sources that will not share data is to run existing data mining tools at each site independently and combine the results[5, 6, 18]. However, this will often fail to give globally valid results. Issues that cause a disparity between local and global results include: Values for a single entity may be split across sources. Data mining at individual sites will be unable to detect cross site correlations. The same item ....

....al. proposed a method for horizontally partitioned data[8] and more recent work has addressed privacy in this model[14] Distributed classification has also been addressed. A meta learning approach has been developed that uses classifiers trained at di#erent sites to develop a global classifier [5, 6, 18]. This could protect the individual entities, but it remains to be shown that the individual classifiers do not disclose private information. Recent work has addressed classification using Bayesian Networks in vertically partitioned data [7] and situations where the distribution is itself ....

P. Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

....procedure m times (once for each set L i ) we obtain the 5 whole meta level data set. Finally, the learning algorithm AML is applied to it in order to induce the combiner CML . The framework for combining multiple classi ers used in this paper is based on the combiner methodology described in [6] and the stacking framework of [22] In these two approaches, only the class values predicted by the base level classi ers are used as meta level attributes. Therefore, the meta level attributes procedure used in these frameworks is trivial: it returns only the class values predicted by the ....

....maximal probability and entropy. This makes them domain dependent in the sense discussed in Section 3. The indirect use of class probability distributions through their properties makes MDTs domain independent. Ordinary decision trees have already been used for combining multiple classi ers in [6]. However, the emphasis of their study is more on partitioning techniques for massive data sets and combining multiple classi ers trained on di erent subsets of massive data sets. Our study focuses on combining multiple classi ers generated on the same data set. Therefore, the obtained results ....

Chan, P. K. and Stolfo, S. J. (1997) On the Accuracy of Meta-learning for Scalable Data Mining. Journal of Intelligent Information Systems 8(1): 5-28. 29

....used in this study are naive Bayes and Linear Discriminant. The integration of base level classi ers is much tighter than in stacking. The similarity to our approach is that class probability distributions are used. Ordinary decision trees have already been used for combining multiple models in [3]. However, the emphasis of their study is more on partitioning techniques for massive data sets and combining multiple models trained on di erent subsets of massive data sets. Our study focuses on combining multiple models generated on the same data set. Therefore, the obtained results are not ....

Chan, P. K. and Stolfo, S. J. (1997) On the Accuracy of Meta-learning for Scalable Data Mining. Journal of Intelligent Information Systems 8(1): 5-28.

....et al. proposed a method for horizontally partitioned data[CNFF96] this is basically the approach outlined in the Figure 2. Distributed classification has also been addressed. A meta learning approach has been developed that uses classifiers trained at individual to develop a global classifier [Cha96, Cha97, PCS00]. This could protect the individual entities, but it remains to be shown that the individual classifiers do not release private information. Recent work has addressed classification in vertically partitioned 3 data [CSK01] and situations where the distribution is itself interesting with respect ....

Philip Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

....modifications of standard learning algorithms, such as decision trees [12] and rule learner [4] An alternative and fairly general method for distributed learning is to combine different multiple predictors in a black box manner. Different meta learning techniques explored at the Jam project [3] were proposed in order to coalesce the predictions of classifiers trained from different partitions of the training set. Similarly, a knowledge probing approach [7] for distributed learning from homogeneous data sites in the first phase learns a set of base classifiers in parallel, and in the ....

Chan, P. and Stolfo, S. On the Accuracy of Meta-learning for Scalable Data Mining, Journal of Intelligent Integration of Information, (Kerschberg L. Ed.), (1998).

....been proposed for distributed data mining. Cheung et al. proposed a method for horizontally partitioned data[CNFF96] Distributed classification has also been addressed. A meta learning approach has been developed that uses classifiers trained at individual to develop a global classifier [Cha96] [Cha97], PCS00] This could protect the individual entities, but it remains to be shown that the individual classifiers do not release private information. Recent work has addressed classification using Bayesian Networks in vertically partitioned data [CSK01] and situations where the distribution is ....

Philip Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

....from changes in the target concept much faster than methods that use all of the training points in a single model. The next step of this research is to parallelize the algorithm in the straightforward way and compare results, in terms of both speedup and accuracy, to other meta learning methods [7, 13]. In the long term, we will focus on improving the generalization accuracy of the method. While the accuracy appears to be about the same as a single classi er, our use of ensemble classi cation suggests that we can do signi cantly better. One direction is to examine the diversity mechanism. We ....

P. K. Chan and S. J. Stolfo. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5-28, 1997.

....this procedure m times (once for each set L i ) we obtain the whole meta level data set. Finally, the learning algorithm AML is applied to it in order to induce the combiner CML . The framework for combining multiple classi ers used in this paper is based on the combiner methodology described in [5] and the stacking framework of [22] In these two 5 approaches, only the class values predicted by the base level classi ers are used as metalevel attributes. In our approach, we also use the class probability distributions predicted by the base level classi ers. The meta level attributes used in ....

....as maximal probability and entropy. This makes them domain dependent in the sense discussed in Section 3. The indirect use of class probability distributions through their properties makes MDTs domain independent. Ordinary decision trees have already been used for combining multiple classi ers in [5]. However, the emphasis of their study is more on partitioning techniques for massive data sets and combining multiple classi ers trained on di erent subsets of massive data sets. Our study focuses on combining multiple classi ers generated on the same data set. Therefore, the obtained results are ....

Chan, P. K. and Stolfo, S. J. (1997) On the Accuracy of Meta-learning for Scalable Data Mining. Journal of Intelligent Information Systems 8(1): 5-28.

....selection of bias takes place. The dominant (task) region for the meta learner may be di erent from the base learners, but ultimately xed (Section 3 and Figure 1) Research in the stacked generalization paradigm investigates what base learners and meta learners produce best empirical results (Chan and Stolfo, 1998; Chan and Stolfo, 1993; Chan, 1996) After transforming the original training set, each example contains the predictions of the base learners, but it may also contain the original features. Results show how certain combinations of learners and meta learners can yield signi cant improvements in ....

....but it may also contain the original features. Results show how certain combinations of learners and meta learners can yield signi cant improvements in accuracy. metaL.tex; 27 03 2001; 11:59; p.7 8 Vilalta and Drissi Several variations to stacked generalization have been explored. For example, Chan and Stolfo (1998) experiment with a modi ed approach where each base learner is trained with a fraction of the total data. While running each learning algorithm in parallel, a hierarchical tree structure is built where each leaf is a level 0 generalizer and each internal node is a high level generalizer (see ....

Chan Philip K. and Stolfo S. (1998). On the Accuracy of Meta-Learning for Scalable Data Mining. Journal of Intelligent Integration of Information, L. Kerschberg, Ed.

....not only at the example (i.e. base) level, but also at the across task (i.e. meta) level. Despite the promising research direction offered by meta learning, no apparent consensus exists of what is meant by such term. Examples of di erent views abound: building a meta learner of base learners [2], selecting inductive biases dynamically [3] building metarules matching task properties with algorithm performance [4] inductive transfer and learning to learn [8] learning classi er systems [5] etc. After addressing some common views of meta learning, this paper proposes a perspective view ....

Chan Philip and Stolfo S. On The Accuracy Of Meta-Learning For Scalable Data Mining. Journal of Intelligent Integration of Information, 1998.

.... one learning algorithm to train different classifiers on subsamples of the training set and then voting to combine the classifications of those classifiers [4,18] Two effective classifiers combination strategies based on stacked generalization (called an arbiter and combiner) were analyzed in [5] showing experimentally that the hierarchical combination approach based on the use of the subsets of the dataset is able to sustain the same level of accuracy as a global classifier trained on the entire data set. One of the most popular and simplest selection approaches is CVM (CrossValidation ....

Chan, P., Stolfo, S.: On the Accuracy of Meta-Learning for Scalable Data Mining. Intelligent Information Systems, Vol. 8 (1997) 5-28.

....meta learning could guide exploration by supplying priorities for the allocation of limited resources, hopefully in a decisiontheoretically sound way. As to guide such a selection or ranking of learning algorithms, several approaches to meta learning have been proposed (e.g. see (Bensusan, 1999; Chan Stolfo, 1996; Giraud Carrier Hilario, 1998; Giraud Carrier Pfahringer, 1999; Lindner Studer, 1999; Widmer, 1997) Perhaps, the most popular strategy consists of describing learning tasks in terms of a set of (meta )attributes and classifying them according to the performance of one or more learners ....

Chan, P., & Stolfo, S. (1996). On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8, 3--28.

No context found.

Chan, Philip and Salvatore Stolfo. 1997. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8(1):5--28.

No context found.

Chan P.K. and Stolfo S.J, On the Accuracy of Meta-learning for Scalable Data Mining, J. Intelligent Information Systems, 8:5-28, 1997.

No context found.

Chan, P., & Stolfo, S. #1996#. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8, 3#28.

No context found.

Chan P. K., Stolfo S. J., On the Accuracy of Meta-Learning for Scalable Data Mining. Journal of Intelligent Information Systems 8(1): 5-28, (1997).

No context found.

Philip Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

No context found.

Chan P. K., Stolfo S. J., On the Accuracy of Meta-Learning for Scalable Data Mining. Journal of Intelligent Information Systems 8(1): 5-28, (1997).

No context found.

P. Chan, S. Stolfo, On the accuracy of meta-learning for scalable data mining. Intelligent Information Systems 8 (1997) 5-28.

No context found.

Philip Chan. On the accuracy of meta-learning for scalable data mining. Journal of Intelligent Information Systems, 8:5--28, 1997.

No context found.

P. Chan and S. J. Stolfo. On the Accuracy of Meta-learning for Scalable Data Mining. Intelligent Information System, 8:5--28, 1996.

No context found.

Chan P, Stolfo S. On the accuracy of meta-learning for scalable data mining. J Intelligent Integration of Information 1997; 8(1):8--25

No context found.

Chan, P., Stolfo, S.: On the accuracy of meta-learning for scalable data mining, Intelligent Information Systems, 8, 1997, 5--28.

First 50 documents

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC