Chan, P. C., & Stolfo, S. (1993). Meta-learning for multistrategy and parallel learning. Proceedings of the Second International Workshop on Multistrategy Learning.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....It has been applied with success to a number of applications, playing an important role to new research areas like Distributed Data Mining, Multiple Classifier Systems and Information Fusion. A particular approach that has been successfully applied to Distributed Data Mining is Meta Learning [2]. It is based on the methodology of Stacked General ization, also found as Stacking in the literature, originally developed by Wolpert [9] Meta Learning involves learning a global classifier that models the way that the output of many local classifiers correlates with the true class. The ....

....for learning the GC is concerned, Ting and Witten [8] showed that among a decision tree learning algorithm (C4.5) an instance based learning algorithm variant (IB1) a multi response linear regression algorithm (MLR) and a Naive Bayes classifier, MLR had the best performance. Chan and Stolfo [2], applied the concept of Stacked Generalization to dis tributed data mining, via their Meta Learning methodology. They focused on combining distributed disjoint data sets and investigated various schemes for structuring the meta level training examples. They found the best one to be using the ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proceedings of the Second International Workshop on Multistrategy Learning, 1993.

....1989 ] determine the weights by assigning, initially, the same value to all classifiers and by decreasing the weights of the wrong classifiers when the collective prediction is false. 2.3. 2 Stacking The main di#erence between voting and stacking [ Wolpert, 1992 ] also called classcombiner [ Chan Stolfo, 1993a ] is that the latter combines base classifiers in a non linear fashion. The combining task, called a meta learner, integrates the independently computed base classifiers into a higher level classifier, the meta classifier, by learning over the meta level training set. This meta level training ....

....It is worth noting, here, that this final classification may be entirely di#erent from those of the constituent base classifiers. Other meta learning approaches that can be considered as variations of the classcombiner strategy, include the class attribute combiner and the binary class combiner [ Chan Stolfo, 1993a ] Again, the goal is to learn the characteristics and performance of the base classifiers and compute a meta classifier model of the global data set. The two combiner strategies di#er from class combiner in that they adopt di#erent policies to compose their meta level training sets. The former ....

Chan, P., and Stolfo, S. 1993a. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, 150--165.

....based learning algorithm variant (IB1) a multi response linear regression algorithm (MLR) and a Naive Bayes classifier, MLR had the best performance. Experimental results from the same work showed that on average Stacking is better than Majority Voting in terms of accuracy. Chan and Stolfo [2], applied the concept of Stacked Generalization to distributed data mining, via their Meta Learning methodology. They focused on combining distributed data sets and investigated various schemes for structuring the meta level training examples. They showed that Meta Learning exhibits better ....

Philip Chan and Salvatore Stolfo, `Meta-learning for multistrategy and parallel learning', in Proceedings of the Second International Workshop on Multistrategy Learning, (1993).

....and Median. 7] is an interesting study of these rules. Stacked Generalization [13] also known as Stacking in the literature, is a method that combines multiple classifiers by learning the way that their output correlates with the true class on an independent set of instances. Chan and Stolfo [2], applied the concept of Stacked Generalization to distributed data mining, via their Meta Learning methodology. They focused on combining distributed data sets and investigated various schemes for structuring the meta level training examples. Knowledge Probing [5] builds on Meta Learning and in ....

Philip Chan and Salvatore Stolfo. Meta-learning for multistrategy and parallel learning. In Proceedings of the Second International Workshop on Multistrategy Learning, 1993.

....relatively new field of Distributed Data Mining (Kargupta Chan, 2000) These include both new algorithms that allow for distributed learning, as well as new system architectures that support the actual learning process. One of the most promising lines of research in this field is Meta Learning (Chan Stolfo, 1993), which is a methodology for deriving a single global classification model by learning from multiple local classifiers. A problem with Meta Learning and other approaches of this field is that they do not take into account the uncertainty and ambiguity involved in the distributed learning ....

....learning can also lead to an increase in accuracy, especially when different learning algorithms with different biases are used. This is due to the fact that each algorithm may compensate the inefficiencies of the other. A very attractive method for distributed learning is Meta Learning (Chan Stolfo, 1993). This approach is based on combining the output of several classification systems (base classifiers) into a final classifier. The combination involves learning the actual way that the output of several classifiers correlates with the true class given a training example. There are two approaches ....

Chan, P., & Stolfo, S. (1993). Meta-Learning for multistrategy and parallel learning. Second International Workshop on Multistrategy Learning.

....techniques generally limit them to classifying small datasets. The datasets for data mining applications are large and may involve several million records. Further, each record typically consists of ten to hundreds attributes. Using large datasets usually improves the accuracy of the classi er [3, 5], but the enormity and complexity of the data involved in these applications makes the classi cation task computationally intensive. Since datasets are large, they cannot reside completely in the main memory, which makes I O a signi cant bottleneck. Performing classi cation for such large datasets ....

.... Code C1 C1 C2 C1 C2 C1 C1 C2 C1 5 6 7 8 9 4 3 2 1 30 25 21 43 18 33 29 55 48 12033 12033 23409 12033 23409 23409 23409 12033 12033 The Training Set Categorical C1 C1 C2 C2 A Decision Tree (Zip Code = 12033) No Yes Yes Yes No No (Age = 48) Age = 21) [1,2,3,4,5,6,7,8,9] [3,5,6,7] A splitting criteria Records of the right sub tree [1,2,4,9] 8] 1,2,4,8,9] 3,5] 6,7] Numeric ( Zip Code = 12033) AND (Age = 48) Description of class C1: Zip Code = 12033) AND (Age 21) OR Figure 2: An example of a decision tree function Decision Tree Classi ....

[Article contains additional citation context not shown here]

P. K. Chan and S. J. Stolfo. Meta-Learning for Multistrategy and Parallel Learning. Proc. Second Int'l Workshop on Multistrategy Learning, pages 150-165, May 1993.

....[Qui93, FI93, Maa94, DKS95] Catlett [Cat91] proposed sampling at each node of the classification tree, but considers in his studies only datasets that could fit in main memory. Methods for partitioning the dataset such that each subset fits in main memory are considered by Chan and Stolfo [CS93a, CS93b] although this method enables classification of large datasets their studies show that the quality of the resulting decision tree is worse than that of a classifier that was constructed taking the complete database into account at once. In this paper, we present a framework for scaling up ....

P. K. Chan and S. J. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Workshop on Multistrategy Learning, 1993.

.... use of larger datasets improves the classification accuracy even further[2] Recently proposed classifiers SLIQ [7] and SPRINT [10] use entire dataset for classification and are shown to be more accurate as compared to the classifiers that use sampled dataset or multiple partitions of the dataset [2, 3]. The decision tree model is built by recursively splitting the training set based on a locally optimal criterion until all or most of the records belonging to each of the partitions bear the same class label. Briefly, there are two phases to this process at each node of the decision tree. First ....

P. K. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. of Second International Workshop on Multistrategy Learning, pages 150--165, 1993.

....have been proposed first in data mining studies. Two strategies for inducing a classification tree from large databases include discretizing continuous attributes and sampling data at each node [6] These, however, still assume that the training set can fit in memory. A method was proposed in [7, 8] that first partitions the data into subsets which individually can fit into memory, and then builds a classification tree from each subset. The final 10 CHAPTER 2. RELATED WORK 11 output classifier combines all the classifiers obtained from the subsets. Although this method can handle large ....

P. K. Chan and S. J. Stolfo. Metalearning for multistrategy and parallel learning. In Proc. 2nd. Int. Workshop on Multistrategy Learning, pages 150--165, 1993.

....Department of Computer Science, University of Rochester, Rochester, NY 14627, zaki cs.rochester.edu. 1 data mining. Developing classification models using larger training sets can enable the development of higher accuracy models. Various studies have confirmed this hypothesis [5] 6] [7]. Examples of recent classification systems that can handle disk resident data include SLIQ [14] and its extension SPRINT [18] A continuing trend in data management is the rapid and inexorable growth in the data that is being collected. The development of high performance scalable data mining ....

Philip K. Chan and Salvatore J. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Workshop on Multistrategy Learning, pages 150--165, 1993.

.... based classifiers that can handle large data sets are important because use of larger data sets improves the classification accuracy [ Michie et al. 1994 ] Previous work in classifying large data sets has been to use sampled data sets or multiple partitions of the data set [ Michie et al. 1994, Chan and Stolfo, 1993 ] Recent work has focused on using the entire data set, in classifiers like SLIQ [ Mehta et al. 1996 ] and SPRINT [ Shafer et al. 1996 ] A parallel classifier in the same spirit has been developed in ScalParC [ Joshi et al. 1998 ] Classifiers like CART [ Breiman et al. 1984 ] and C4.5 [ ....

P.K Chan and S.J Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. International Workshop on Multistrategy Learning, 1993.

....nor feasible, to inspect all of the data at one processing site to compute one primary global classifier or model at a reasonable cost. We call the problem of learning useful new information from large and inherently distributed databases, the scaling problem for machine learning. Meta learning [8], a technique similar to stacking [40] was developed recently to deal with the scaling problem. The basic idea is to execute a number of machine learning processes on a number of data subsets in parallel, and then to combine their collective results (classifiers) through an additional phase of ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....final sorted list of ranked pages. Using different learning algorithms and combining the learned models has been demonstrated to be effective in improving performance [3, 4, 8, 37, 46] We plan to investigate using multiple estimators to analyze and combine the results in parallel. Meta learning [6] is one combining approach and we intend to examine other voting based methods [7] Properly evaluating how the user responds to the results returned by our search engine is important. In information retrieval, the common metrics are precision and recall. From all pages returned by the search ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....We address the efficiency and scalability issues in several ways. We divide a large data set of labeled transactions (either fraudulent or legitimate) into smaller subsets, apply mining techniques to generate classifiers in parallel, and combine the resultant base models by meta learning [1] from the classifiers behavior to generate a meta classifier. Our approach treats the classifiers as black boxes so that a variety of learning algorithms can be employed. Besides extensibility, combining multiple models computed over all available data produces meta classifiers that can ....

....process can be run in parallel. For massive amounts of data, substantial improvement in speed can be achieved for super linear time learning algorithms. The generated classifiers are combined by meta learning from their classification behavior. Several meta learning strategies are described in [1]. To simplify our discussion, we only describe the class combiner (or stacking [12] strategy. In this strategy a meta level training set is composed by using the (base) classifiers predictions on a validation set as attribute values and the actual classification as the class label. This training ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....We address the efficiency and scalability issues in several ways. We divide a large data set of labeled transactions (either fraudulent or legitimate) into smaller subsets, apply mining techniques to generate classifiers in parallel, and combine the resultant base models by meta learning [1] from the classifiers behavior to generate a meta classifier. Our approach treats the classifiers as black boxes so that a variety of learning algorithms can be employed. Besides extensibility, combining multiple models computed over all available data produces meta classifiers that can ....

....process can be run in parallel. For massive amounts of data, substantial improvement in speed can be achieved for super linear time learning algorithms. The generated classifiers are combined by meta learning from their classification behavior. Several meta learning strategies are described in [1]. To simplify our discussion, we only describe the class combiner (or stacking [12] strategy. In this strategy a meta level training set is composed by using the base classifiers predictions on a validation set as attribute values and the actual classification as the class label. This training ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....value. In this study, the weights w i s are set according to the performance (with respect to a selected evaluation metric, e.g accuracy) of each classifier C i on a separate validation set. 2. 2 Stacking The main di#erence between voting and Stacking [40] or Class Combiner Meta learning [5]) is that the latter combines base classifiers in a non linear fashion. The combining task, called a meta learner, integrates the independently computed base classifiers into a higher level meta classifier, by learning over a meta level training set. This meta level training set is composed by ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....be possible to inspect all of the data at one processing site to compute one primary global classifier or model at a reasonable cost. We call the problem of learning useful new information from large and inherently distributed databases, the scaling problem for machine learning. Meta learning [7], a technique similar to stacking [26] was developed recently to deal with the scaling problem. The basic idea is to execute a number of machine learning processes on a number of data subsets in parallel, and Supported in part by an IBM fellowship This research is supported by the Intrusion ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. 2nd Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....distributions used by Bayesian classifiers are derived from the frequency distributions of attribute values and reflect the likelihood of a certain instance belonging to a particular classification [12] Implicit decision rules classify according to maximal probabilities. Meta learning [7] is loosely defined as learning from learned knowledge. In this case, we concentrate on learning from the output of concept learning systems. This is achieved by learning from the predictions of these classifiers on a common validation data set. Thus, we are interested in the output of the ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....using a variety of different machine learning algorithms (L j ) in order to determine the best strategies for building good fraud detectors. The best base classifiers are then combined by a variety of techniques in order to boost performance. One of the simplest combining algorithms proposed in[3], and independently by Wolpert[29] is called class combiner or stacking . A separate hold out training dataset, V , is used to generate a meta level training data to learn a new meta classifier M . M is computed by learning a classifier from training data composed of the predictions of a set ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

....of overhead to nonfrauds. This reflects how the prediction errors will add to the total cost of a hypothesis. Because the actual overhead is a closely guarded trade secret and is unknown to us, we chose to set overhead 60, 70, 80, 90 to run four sets of experiments. We normalized each c i to [0,1]. The cost adjustment function b is chosen as: b (c) 0.5c 0.5 and b (c) 0.5c 0.5. As in previous experiments, we use training data from one month and data from two months later for testing. Our data set let us form 10 such pairs of training and test sets. We ran both AdaBoost ....

P. Chan and S. Stolfo, "Metalearning for Multistrategy and Parallel Learning," Proc. Second Int'l Workshop Multistrategy Learning, Center for Artificial Intelligence, George Mason Univ., Fairfax, Va., 1993, pp. 150--165.

....using a variety of different machine learning algorithms (L j ) in order to determine the best strategies for building good fraud detectors. The best base classifiers are then combined by a variety of techniques in order to boost performance. One of the simplest combining algorithms proposed in[3], and independently by Wolpert[29] is called class combiner or stacking . A separate hold out training dataset, V , is used to generate a meta level training data to learn a new meta classifier , M . M is computed by learning a classifier from training data composed of the predictions of a ....

P. Chan and S. Stolfo. Meta-learning for multistrategy and parallel learning. In Proc. Second Intl. Work. Multistrategy Learning, pages 150--165, 1993.

No context found.

Chan, P. C., & Stolfo, S. (1993). Meta-learning for multistrategy and parallel learning. Proceedings of the Second International Workshop on Multistrategy Learning.

No context found.

P. K. Chan and S. Stolfo. MetaLearning for Multistrategy and Parallel Learning. In Proceedings of the Second International Workshop on Multistrategy Learning, pages 150--165. George Mason University, 1993.

No context found.

Philip Chan and Salvatore Stolfo. Meta-learning for multistrategy and parallel learning. In Proceedings of the Second International Workshop on Multistrategy Learning, 1993.

No context found.

Chan, P., & Stolfo, S. (1993a). Meta-learning for multistrategy and parallel learning. Proceedings of the Second International Workshop on Multistrategy Learning (pp. 150--165). Fairfax, VA: Center for AI, George Masion University.

First 50 documents  Next 50

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