Widmer, Gerhard. 1997. Tracking context changes through meta-learning. Machine Learning, 27(3):259--286.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....[4] market basket analysis [6] e mail importance ranking [2] calendar scheduling [7] and intelligent user interfaces [8] Researchers have investigated several methods of retaining and then discarding examples of old concepts. One is to remove examples after a fixed period of time [1] 4] [9]. Another is to determine this period dynamically in response to changing performance [3] We have investigated learners that maintain examples that lie on rule boundaries [2] 4] 10] When new examples arrive, the learner produces new rules and forgets those examples that no longer fall on ....

....boundaries of concepts, IB2 stores examples near this interface. The original motivation for this algorithm was not concept drift, so if used for this purpose, additional mechanisms will be required to remove outdated examples. A simple method is to store examples over a fixed window of time [4] [9]. Naturally, performance is dependent on selecting the correct size of the window, so some researchers have examined heuristics for dynamically sizing these windows [3] Similarly, AQ11 PM has partial instance memory, but it selects examples from the boundaries of its rules [10] An extension of ....

G. Widmer, "Tracking context changes through meta-learning," Machine Learning, vol. 27, pp. 259--286, 1997.

....Markov models for predicting users requests on a WWW server [19] etc. In both cases hybrid models are presented that perform better than any of the original models alone. Finally, some papers propose multilevel learning in order to learn user s interest that can possibly drift and recur [8, 17]. 2 Hierarchy of Beliefs An autonomous agent would obviously be interested in possessing an actual and adequate model of the actual user. It may include the user s preferences, his actual strategy, predicted future actions etc. However, such a model can hardly be acquired: the user may ....

G. Widmer. Tracking context changes through meta-learning. Machine Learning, 27:256--286, 1997.

....there is a hidden unexpected change in context; to apply learning gained in one context to different context; and to utilise already known information about contexts to improve learning. There are only a few papers which touch on the problem of learning the appropriate contexts themselves. Widmer [30] applies a metalearning process to a basic incremental learning neural net; the meta algorithm adjusts the window over which the basic learning process works. Here it is an assumption that contexts are contiguous in time and so a time window is a sufficient representation of context. Harries et ....

Widmer, G. Tracking Context Changes through MetaLearning. Machine Learning, 27. 259-286.

....two descriptors, positive and negative, to keep up with short term interests. This approach is similar to an incremental method for learning in domains with concept drift, where multiple concept representations that generalize examples over different window sizes are maintained simultaneously [14, 15]. Compared to systems that mainly use a single descriptor model for interest category representation, the 3 descriptor scheme has several advantages. The 3 descriptor scheme allows learning of long term and short term interests simultaneously, and also handles exceptions of intexests within an ....

Widmer, G. 1997 Tracking Context Changes through Meta-Learning. Machine Learning, 27(3):259-286, Kluwer Academic Publisher.

....the conditions which the context is recognised) can change more quickly than the membership of that context. In this way I have allowed the emergence of context like behaviour without imposing the sort of two tier structures that have been employed in other machine learning algorithms (e.g. [16, 20]) If I had a network with more than one intermediate layer we could allow the emergence of contexts within contexts etc. but this would take longer runs in order for any results to emerge. AAAA AAAA AAAA AAAA AAAAAAAA AAAAAAAA AAAA AAAAAAAA AAAAAAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA AAAA ....

Widmer, G. (1997). Tracking Context Changes through Meta-Learning. Machine Learning, 27:259-286.

....two descriptors, positive and negative, to keep up with short term interests. This approach is similar to an incremental method for learning in domains with concept drift, where multiple concept representations that generalize examples over different window sizes are maintained simultaneously [14, 15]. Compared to systems that mainly use a single descriptor model for interest category representation, the 3 descriptor scheme has several advantages. The 3 descriptor scheme allows learning of long term and short term interests simultaneously, and also handles exceptions of interests within an ....

Widmer, G. 1997 Tracking Context Changes through Meta-Learning. Machine Learning, 27(3):259-286, Kluwer Academic Publisher.

....are those systems that exploit partial instance memory. HILLARY [14] maintains recent negative examples. FLORA2, FLORA3 and FLORA4 [38] store a consecutive sequence of examples in a certain time window. The size of the time window is adapted permanently. In the opposite, MetaL(B) and MetaL(IB) [37] keep training examples over a fixed window of time. AQ PM [25] maintains extreme instances that lie at the boundaries of the current concept description. The approach presented here combines different ideas. We a priori fix the size of the partial instance memory. The examples stored there are ....

G. Widmer. Tracking context changes through meta-learning. Machine Learning, 27:259--286, 1997.

....in the same descriptor for the long term interest model. This approach is similar to an incremental method for learning in domains with concept drift, where multiple concept representations that generalize examples over different window sizes are maintained simultaneously (Widmer Kubat, 1996; Widmer, 1997). Several fundamental techniques from information retrieval are employed to develop the learning algorithm. Furthermore, the algorithm is designed to work in a setting where other knowledge (such as word document frequency information) is not available at the time the system is first used, and ....

Widmer, G. (1997). Tracking Context Changes through Meta-Learning. Machine Learning Journal, 3, 259-286. Boston, MA: Kluwer Academic Publisher.

....begin to converge toward the target concepts, the size of the window increases, as does the number of training examples maintained in partial memory. Other examples of partial memory learning systems include HILLARY (Iba, Woogulis, Langley, 1988) DARLING (Salganicoff, 1993) and MetaL(B) (Widmer, 1997). Partial Memory Learning 3 0. Learn Partial Memory(Data t , for t = 1 : n) 1. Concepts0 = PartialMemory0 = 2. for t = 1 to n do 3. Missed t = Find Missed Examples(Concepts t Gamma1 , Data t ) 4. TrainingSet t = PartialMemory t Gamma1 [ Missed t ; 5. Concepts t = Learn(TrainingSet ....

Widmer, G. (1997). Tracking context changes through meta-learning. Machine Learning, 27, 259--286.

....or to derive new concept descriptions. Researchers have developed partial memory systems because they can be less susceptible to overtraining when learning concepts that change or drift, as compared to learners that use other memory models (Salganicoff, 1993; Maloof, 1996; Widmer Kubat, 1996; Widmer, 1997). The key issues for partial memory learning systems are how they select the most relevant examples from the input stream, maintain them, and use them in future learning episodes. These decisions affect the system s predictive accuracy, memory requirements, and ability to cope with changing ....

.... Systems with partial instance memory appear to be the least studied, but examples include LAIR (Elio Watanabe, 1991) HILLARY (Iba, Woogulis, Langley, 1988) IB2 (Aha et al. 1991) DARLING (Salganicoff, 1993) AQ PM (Maloof Michalski, 1995) FLORA2 (Widmer Kubat, 1996) and MetaL(B) (Widmer, 1997). On line learning systems must also have policies that deal with concept memory, which refers to the store of concept descriptions. Researchers have investigated a variety of strategies in conjunction with different models of instance memory. For example, IB1 (Aha et al. 1991) maintains all ....

[Article contains additional citation context not shown here]

Widmer, G. (1997). Tracking context changes through meta-learning. Machine Learning, 27, 259--286.

....old contexts and corresponding concepts reappear. They reported a series of experiments with different implementations of FLORA, that has the capability of dynamic windowing search to take into account the effects of concept drift and hidden contexts. Within the scope of meta learning, Widmer [8] introduces a general two level learning model, and the results of implementing the approach in two systems that can learn to detect such contextual clues, and can react accordingly when a change in context is suspected. In this approach a base level learner performs the regular on line learning ....

Widmer, G.: Tracking Context Changes through Meta-Learning. Machine Learning 27 (1997) 259-286.

....of the calendar scheduling problem, the prediction of meeting duration, for a single user, Tom Mitchell. The data set 1 runs from 3 March 1992 to 16 December 1993 and has 1685 entries, each entry has 14 attributes and a classification. These data were previously used by (Mitchell et al. 1994; Widmer 1997). This preliminary result shows the context boundaries found by Splice 2.1 on this data set. Splice 2.1 was run ten times with four randomly drawn initial clusters, a window size of 100 items, and for 20 iterations. The window size was chosen to make the recognition of large scale contexts, such ....

Widmer, G. 1997. Tracking context changes through meta-learning. Machine Learning 27:259--286.

....class membership to find a proper weight vector it can then accurately classify new documents. Such algorithms are called learning classification algorithms. These algorithms will be discussed is section two of this paper. This paper addresses the Simple Bayesian Classifier ( IRFT98] OPSBC96] [CCML96]) How the Simple Bayesian Classifier works, what it needs to perform the classification and so on. 2 Learning classification algorithms Learning classification algorithms use training data to find a weight vector that accurately classifies new texts. Learning classification algorithms can be ....

....Thus the SBC algorithm is the most accurate of the six algorithms for that particular context. It also becomes clear that the SBC algorithm is the most efficient of the six algorithms. For an elaborate description of the conducted experiments the reader is referred to the [IWWW95] article. In the [CCML96] article another example is given how the SBC algorithm can be used in practical situations. This article describes how context changes can be tracked through meta learning. That means that the article deals with the problem of learning incrementally in domains where the target concepts are ....

[Article contains additional citation context not shown here]

G. Widmer. Tracking Context Changes through Meta-Learning. 1996.

No context found.

Widmer, Gerhard. 1997. Tracking context changes through meta-learning. Machine Learning, 27(3):259--286.

No context found.

Widmer, G. (1997). Tracking context changes through metalearning.

No context found.

G. Widmer. Tracking context changes through meta-learning. Machine Learning, 27(3):259-- 286, 1997.

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