GENNARY J., LANGLEY P., FISHER D. 1989. Model of Incremental Concept Formation. Artificial Intelligence Journal, Volume 40, 11-61.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....methods (DISC [11] and PBKI [16] derive hierarchical nearness structures sensitive to the distribution of attribute values in the database. Comparable automatic acquisition techniques do not combine both the hierarchical and distribution sensitive concepts. For instance, COBWEB [19] and CLASSIT [20] use category utility as a similarity measure and do not adjust to the distribution (or population) of data values. The similarity measure used in WITT [21] is based on pair wise attribute correlation rather than the nearness of attribute values and thus it is not suitable for deriving approximate ....

J. Gennari, P. Langley, and D. Fisher, \Models of incremental concept formation," Arti cial Intelligence, vol. 40, pp. 11-62, 1989.

....of the categorical attribute values Vp. Being incremental, COBWEB is fast with a complexity of O(tN) though it depends non linearly on tree characteristics packed into constant t. The similar incremental hierarchical algorithm for all numerical attributes, CLASSIT, was developed by Gennari et al. [GLF89]. CLASSIT associates normal distributions with cluster nodes. Both algorithms can result in a highly unbalanced trees. Chiu et al. CFCW01] proposed another conceptual or model based approach to hierarchical clustering. This development contains several different useful features, such as the ....

Gennari, J., Langley, P., and Fisher, D. Models of incremental concept formation. Artificial Intelligence, 40, 11-61, 1989. 45

....tends to rise as the threshold value increases; however, the increase is not very dramatic, supporting the suitability 12 of L SEuS for preliminary exploratory data analysis. 4 Related Work Much of the prior work on structure discovery is domain dependent (e.g. Win75, Lev84, Fis87, Leb87, GLF89, CG92] and a detailed comparison of these methods appears in [Con94] We consider only domain independent methods in this paper. The first such system, CLIP, discovers patterns in graphs by expanding and combining patterns discovered in previous iterations [YMI93] To guide the search, CLIP uses ....

J. H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, (40):11--61, 1989.

....represent these clusters intensionally (as inductive hypotheses made by abstractions from specific examples) The clustering process is based on some quality criteria generally minimizing the overlapping between different clusters. The clusters are often represented probabilistically (as in COBWEB [5] and AutoClass [3] This approach is called conceptual clustering. 3. The training examples are preclassified and the learning system is simply memorizing them. Then the predictions are made on the basis of the similarity of the new examples to the memorized ones. A typical instance of this ....

J. H. Gennari, P. Langley, and D. Fisher. Model of incremental concept formation. In J. G. Carbonell, editor, Machine Learinng: paradigms and methods. MIT Press, 1990.

.... been discussed in the computational learning theory community (cf. e.g. Wiehagen [29] Kinber and Stephan [13] Lange and Zeugmann [19] Jain et al. 11] Case et al. 4] Lange and Grieser [16] Lange [15] as well as in the machine learning community (cf. e.g. Utgoff [26] Gennari et al. [7], Porat and Feldmann [22] Godin and Missaoui [8] Maloof and Michalski [21] A prominent and intensively studied example is iterative learning. Here, the learning device (henceforth called iterative learner) is required to produce its actual hypothesis exclusively from its previous one and the ....

Gennari, J.H., Langley, P., and Fisher, D., Models of incremental concept formation, Artificial Intelligence 40, 11--61, 1989.

....objects, we can consider that the worst case time complexity of RUE is O(n 2 m 3 ) 4.1. Experimental results In this section we present a comparison between RGC and a selection of the most representative conceptual algorithms elaborated at this moment: Conceptual KMEANS [9] LC [10] COBWEB [11] and LINNEO [12] They all will characterize an available data set ZOO that contains descriptive information about 101 animals described in terms of 15 Boolean features and one numerical variable [13] The task is to analyze the potentiality of these conceptual algorithms in order to give ....

J.H. Gennari, P. Langley and D. Fisher, "Models of Incremental Concept Formation", In Jaime Carbonell, MIT/Elsevier Machine Learning. Paradigms and Methods, 1990, pp. 11-61.

....adding the most similar instances to a cluster described by the current prototype and containing all structures generalised by the prototype including the seed and the instances added in the generalisation step. This is similar to the method of conceptual clustering (see for example [11] and [16]) A similar procedure can be used to find clusters of structured objects with some common structural pattern in a database by unsupervised learning, i.e. where no classification is given a priori. The second algorithm, called SIG Learner, learns only one class at a time. It generalises a set of ....

J. H. Gennari, P. Langley, and D. Fisher. Models of Incremental Concept Formation. Artificial Intelligence, 40:11 -- 61, 1989.

.... by subsequently adding the most similar instances to a cluster described by the current prototype and containing all structures generalised by the prototype including the seed and the instances added in step (2(b)ii) This is similar to the method of conceptual clustering (see for example [13] and [18]) A similar procedure can be used to find clusters of structured objects with some common structural pattern in a database by unsupervised learning, i.e. where no classification is given a priori. This similarity based method requires the computation of the similarity of all remaining examples ....

J. H. Gennari, P. Langley, and D. Fisher. Models of Incremental Concept Formation. Artificial Intelligence, 40:11 -- 61, 1989.

....data. The task is to find a feature set that provides the highest prediction accuracy. The wrapper method proposes a set of features that are relevant for the identification of the expectation violation. Features are considered relevant if their values vary systemically with category membership [Gennari, Langley and Fisher 1989]. Figure 3: Wrapper based expectation acquisition Metareason. mod. Inductive learner Accur. testing Feat. select. Expect. repres. Features in the observable world of the agent Candidate features for learning Wrapper Figure 3 summarizes the learning process in the context of ....

Gennari, J.H., Langley, P., Fisher, D. Models of incremental concept formation, Artificial Intelligence, 40, 1989, pp. 11-61.

....and the types of conflicts associated with them. Keeping design histories would be one way to achieve this, where situation analysis is done in a case based manner. However, by defining a set of attributes to characterize conflicts, it would be possible to take an inductive classification approach [Gennari et al. 1989]. The knowledge acquired helps anticipate conflicts and base proposals on expectations. Learning negotiation plans: Keeping negotiation histories represents an opportunity to generate plans from previous negotiation cases. This can be done by compiling negotiation sequences using an explanation ....

Gennari et al. 1989 J.H. Gennari, P. Langley & D. Fisher. "Models of Incremental Concept Formation ", Artificial Intelligence, Vol. 40, No. 1-3, 1989, pp.11-61.

....We deal with iterative learning, k bounded examplememory inference, and feedback identification of indexable concept classes. All these models formalize incremental learning, a topic attracting more and more attention in the machine learning community (cf. e.g. Utgoff [32] Gennari et al. [11], Porat and Feldmann [29] Godin and Missaoui [12] Langley [27] Maloof and Michalski [28] An iterative learner is required to produce its actual guess exclusively from its previous one and the next element in the information sequence presented. Iterative learning has been introduced in ....

J. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--62, 1989.

....the unknown example to be classified, the most similar generalized subgraph is used for classification, by using Eq. 2) Clearly MatchBox s concept formation adopts the idea of conceptual clustering introduced by Michalski et al. 73] and developed, for instance, by Lebowitz [74] Fisher et al. [75] and others. Because it is impossible to compute all possible generalizations, many authors use a similarity measure for objects as a guideline for generalization, for example, Lebowitz for attribute value descriptions, Bisson [19, 76] for logical representations, and the distance guided ....

J. H. Gennari, P. Langley, and D. Fisher. Models of Incremental Concept Formation. Artificial Intelligence, 40:11 -- 61, 1989.

.... Dietterich (1991, p. 548) define relevance under the assumptions that all features and the label are Boolean and that there is no noise. Definition 2 A feature X i is said to be relevant to a concept C if X i appears in every Boolean formula that represents C and irrelevant otherwise. Gennari, Langley Fisher (1989, Section 5.5) allow noise and multi valued features and define relevance as 1 Definition 3 X i is relevant iff there exists some x i and y for which p(X i = x i ) 0 such that p(Y = y j X i = x i ) 6= p(Y = y) Under this definition, X i is relevant if knowing its value can change the ....

Gennari, J. H., Langley, P. & Fisher, D. (1989), "Models of incremental concept formation", Artificial Intelligence 40pp. 11--61.

....with mean equal to the value of the data point and standard deviation zero. The idea of using statistical information obtained from a training set in the formation of prototypes is not a novel idea, either. Two examples of similar systems are radial basis function networks [Hay94] and CLASSIT [Gen90]. CLASSIT is an unsupervised concept learning system based on the same model as Fischer s COBWEB [Fis90] Gen90] Both CLASSIT and COBWEB create a hierarchy of concept descriptions in an attempt to discover natural classes in the training data. Basic learning begins with the presentation of an ....

....obtained from a training set in the formation of prototypes is not a novel idea, either. Two examples of similar systems are radial basis function networks [Hay94] and CLASSIT [Gen90] CLASSIT is an unsupervised concept learning system based on the same model as Fischer s COBWEB [Fis90][Gen90]. Both CLASSIT and COBWEB create a hierarchy of concept descriptions in an attempt to discover natural classes in the training data. Basic learning begins with the presentation of an instance to the system, which then chooses from four possible operations: incorporate the instance into a given ....

Gennari, John H.; Langley, Pat; and Fisher, Doug, "Models of Incremental Concept Formation", Machine Learning: Paradigms and Methods, ed. Jaime Carbonell, MIT press, Cambridge, Massachusetts, pp. 11-61, 1990.

....73.3 Avoid Type A Error 91.0 36.0 71.8 Avoid Type B Error 67.9 74.6 70.2 23 8.5 Heart Disease Diagnosis Observations for 303 patients are given, of which 165 are healthy, while 139 have some heart disease. Of the 303 records, 6 have some missing entries. For prior computational results, see Gennari et al. 1989), Shavlik et al. 1991) Holte (1993) and Boros et al. 1996) Each record provides 13 attributes, of which 3 are Boolean, 4 nominal, and 6 rational. We transform the records to logic data and obtain a total of 50 logic variables. We collect in A (resp. B) the logic records corresponding to the ....

Gennari, J. H., Langley, P., and Fisher, D.1989 Models of Incremental Concept Formation. Articial Intelligence 40 11-61.

....to measuring correlation between nominal variables are discussed in Section 4.2; their respective biases and implications for use with CFS are discussed in Section 4.3. Section 4.4 presents the CFS algorithm and the variations used for experimental purposes. 4. 1 Rationale Genari et al. GLF89] state that Features are relevant if their values vary systematically with category membership. 1 The term correlation is used in its general sense in this thesis. It is not intended to refer specifically to classical linear correlation; rather it is used to refer to a degree of dependence ....

J. H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, (40):11--61, 1989.

....or to provide support to create a new one. This observation also applies to classification algorithms. No methodology or tool has been proposed to support the elaboration of conceptual clustering algorithms that build task specific ontologies. Work on conceptual clustering (e.g. 19] 8] [9], 2] 1] 26] has not been extensively applied to the problem of learning from corpora. One must however acknowledge that the application of conceptual clustering techniques to this domain is not straightforward, as existing algorithms must be previously adapted. As in the case of distances, ....

Gennari J., Langley , P., Fisher D. 1989. Model of Incremental Concept Formation, Artificial Intelligence Journal, Volume 40, 11-61

....consistent with specialization. Another advantage of the theory is that it provides a clear and simple definition of the nature of the generated hierarchy that does not depend on algorithm specifics or parameter tuning, as opposed to most work in concept formation methods (Fisher Pazzani, 1991; Gennari, Langley Fisher, 1990) and other algorithms for dealing with class hierarchies (Casais, 1991; Dvorak, 1994; Lieberherr et al. 1991) Because the design of class hierarchies is typically an iterative and incremental process, we developed efficient incremental algorithms that update the hierarchies by adding, removing, ....

Gennari, J. H., Langley, P. & Fisher, D. (1990). Models of Incremental Concept Formation. In J . Carbonell (Eds.), Machine Learning: Paradigms and Methods, (pp. 11-62). Amsterdam, The Netherlands: MIT Press.

....Nouns Disambiguation Lexical Tuning Task Tester Filled Templates Figure 1. The functional architecture for Lexical Tuning 3 Learning subcategorization patterns from corpora Methods of clustering are widely adopted within the machine learning community for example driven learning tasks [J.H. Gennari1989] They are generally based on incremental search within concept (or class) spaces. The target problem here is the derivation of the subcategorization frame for each verb observed in the source training text. In this case a separate search space is used for each verb. In particular: 1. All the ....

P. Langley D.H. Fisher J.H. Gennari. 1989. Models of incremental concept formation. Artificial Intelligence, 40.

....threshold. For an object to belong to a cluster it needs to be similar enough to at least one other member of the cluster. To characterize the similarity between a pair of documents we used the cosine coefficient. 3.2. 2 CLASSIT AGGLOM The most well known conceptual clustering system is COBWEB [11, 17]. It creates clusters that are characterized by the list of nominal attribute values and probabilities associated with them. COBWEB s evaluation function, category utility [11, 14] estimates not the similarity between individual objects, but the overall quality of the partition. In our ....

....by the list of nominal attribute values and probabilities associated with them. COBWEB s evaluation function, category utility [11, 14] estimates not the similarity between individual objects, but the overall quality of the partition. In our experiments, we use the successor to COBWEB (CLASSIT [17]) that extends these ideas onto continuous attribute values. We had to modify CLASSIT s evaluation function to account for the missing terms [13] CU C 1 , C 2 , C K ( P C k ( 1 I k P A C k ( 1 s ik i I k k K 1 I P A i ( 1 s i i I K , where P C k ( ....

Gennari, J. H., Langley, P., and Fisher, D. Models of incremental concept formation. Artificial Intelligence, 40, pp. 11-61; 1989.

....to build organizations of relational concepts. Keywords: Relational data, Unsupervised learning, Reformulation 1 Introduction In Artificial Intelligence, the problem of automatically constructing classifications has been the subject of much research during the last fifteen years [21] 11] [14]. It consists in searching for similarities among objects that are not pre classified and structuring them in a hierarchy in which similar objects are clustered. Most of the existing Conceptual Clustering approaches have defined this task as the search for a classification that would best predict ....

....among objects that are not pre classified and structuring them in a hierarchy in which similar objects are clustered. Most of the existing Conceptual Clustering approaches have defined this task as the search for a classification that would best predict unknown features of new objects [10] 12] [14]. This type of construction is guided by heuristics, which allow one to choose the best concepts among the set of possible ones. The developed methods have proved their interest in various fields [21] 11] 14] 17] More recent research concerns the construction of classifications that ....

[Article contains additional citation context not shown here]

Gennari J. H., Langley P., Fisher D.: Models of incremental concept formation. Artificial Intelligence 40-1(3). (1989). 11-61.

....and we have used Aide to perform preliminary exploration on real data. We have not yet begun a systematic evaluation, however. 2 Themes in Exploration Existing approaches to exploration fall into two classes: autonomous machine discovery systems that perform clustering or function finding [ Gennari et al. 1989; Biswas et al. 1991; Langley et al. 1987; Schaffer, 1990 ] and user driven statistics packages. These approaches occupy opposite endpoints on a spectrum of control. In applying a machine discovery system to a task, the user specifies the goals of the analysis implicitly in the form of the ....

Gennari, John H.; Langley, Pat; and Fisher, Doug 1989. Models of incremental concept formation. Artificial Intelligence 40.

.... The task of concept formation is to take a large number of unlabeled training instances; to find clusterings that group those instances in categories; to find an intensional definition for each category that summarizes its instances; and to find a hierarchical organization for those categories[Gennari et al. 1989]. 2 Related Work Much of the recent research on this topic builds on Fisher s [1987] COBWEB. Fisher s system assumes that each instance is described as a conjunction of attribute value pairs, and employs a probabilistic representation for concepts. All attributes of an instance are used in ....

....those links that contain features in the instance, and it continues the process with the relevant children. Both the number of features necessary for this match and closeness of each value (for numeric attributes) are system parameters. For a detailed review of COBWEB and UNIMEM, please refer to [Gennari et al. 1989]. 3 A Joint Concept Formation System We propose SGNN 1 here as a joint concept formation system which bears many similarities to COBWEB. It represents knowledge into a hierarchy of concepts, classifies new instances by sorting them down this hierarchy. However, SGNN can construct a concept ....

[Article contains additional citation context not shown here]

J.H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--61, 1989.

....this study has taken place in the field of inductive learning. Most of the developed approaches have defined this task as the search for a classification that would best predict unknown features of new objects. The developed methods have proved successful through a number of applications [18] 7] [10]. Our study registers in the framework of the automatic construction of classifications but we tackle an issue which has been less explored, that of the discovery of classifications. In this prospect, the predictive capacity of a classification becomes much less important than its intrinsic power ....

J.H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. In Artificial Intelligence 40 - 1(3), pages 11--61. 1989.

....data instances. Specifically, wrappers eliminate conditions from the candidate condition set. The wrapper method proposes a subset of features that are relevant for the identification of a given class. Features are considered relevant if their values vary systemically with category membership (Gennari, Langley, and Fisher 1989), in our case, with the ranges of the expectation target. For this purpose the wrapper maintains several subsets of candidate features. An accuracy testing component determines the performance of each subset, and eliminates or adds new subsets of features, by providing information to a feature ....

Gennari, J.H., P. Langley, and D. Fisher. (1989). Models of incremental concept formation. Artificial Intelligence 40:11-61.

....a natural solution to cope with such situations because they are able to incorporate into the model new samples from the changing world. 2.3. 1 Incremental algorithms: a definition Incremental learning algorithms have been largely studied in Conceptual Clustering [1, 11, 22] and Concept Formation [13, 2] communities. The idea of incrementality arises from the observation that much of human learning can be viewed as a gradual process of concept formation and human ability for incorporating knowledge from new experiences into already learnt concept structures. The incremental learning approach is ....

....searches and, in this way, explore different optima. However, the algorithms using the beam search are usually provided with parameters which state the number of beams to be used in order to control the amount of time and memory required to perform the search. In addition, it is observed [10, 13, 21, 22, 2] that incremental hill climbing, in the context of unsupervised learning, is order sensitive. Namely, given two sample orders, O 1 and O 2 , of a database D an incremental hill climbing algorithm may output different domain models when fed with order O 1 or with order O 2 . Ordering effects are ....

J. H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, (40):11--61, 1989.

....of an extension representing a subset of instances and an intension representing the common features for this set of instances. The task of inducing a concept hierarchy in an incremental manner is known as incremental concept formation or simply concept formation (Fisher, Pazzani Langley, 1991; Gennari, Langley Fisher, 1990) and is a fundamental process of human learning. In this article we present and analyze incremental algorithms for building Galois lattices. The main distinguishing characteristics of this method with respect to other well known concept formation methods such as UNIMEM (Lebowitz, 1987) COBWEB ....

....of human learning. In this article we present and analyze incremental algorithms for building Galois lattices. The main distinguishing characteristics of this method with respect to other well known concept formation methods such as UNIMEM (Lebowitz, 1987) COBWEB (Fisher, 1987) and CLASSIT (Gennari et al. 1990) are that the concept hierarchy is a lattice, not restricted to a tree hierarchy, and although the algorithm is incremental, the resulting hierarchy does not depend on any tunable parameters, algorithm specifics, or the order in which the instances are acquired. A good overview of work on concept ....

[Article contains additional citation context not shown here]

Gennari, J. H., Langley, P. & Fisher, D. (1990). Models of Incremental Concept Formation. In J. Carbonell (Eds.), Machine Learning: Paradigms and Methods, (pp. 11-62). Amsterdam, The Netherlands: MIT Press.

....does point to the difficulty of the matching process with complex objects. However no claim of handling large amounts of knowledge objects could be made until a detailed complexity analysis is available, or until extensive experimentation with real data is accomplished. The CLASSIT system [8] improves on COBWEB by dealing with numeric attributes and by allowing richer descriptions of instances (defined as sets of elements) However, every element in a set must have the same structure which is used to simplify the matching process. Thus, the representational capabilities of this system ....

J.H. Gennari, P. Langley, and D. Fisher, "Models of Incremental Concept Formation", J. Carbonell, Ed., Machine Learning: Paradigms and Methods, Amsterdam, The Netherlands: MIT Press, 11-62, 1990.

.... descriptions will be enhanced later by results of experimentation (e.g. mass, center of gravity, face normals) Several descriptions of one view of a type of objects are then combined into a generic description by using an algorithm based on ideas of the conceptual clustering algorithm CLASSIT [11]. We extended the algorithm so that it is able to handle relational attributes in the matching step. As the algorithm inserts a new 2Ddescription in the root of the concept tree and then performs an incremental hill climbing search in the tree (together with two restructuring operations) the ....

J.H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--61, 1989.

....two applications of BIRCH, which are also intended to show how BIRCH, CLARANS and KMEANS perform on some real datasets. Finally our conclusions and directions for future research are presented in Section 8. 2. Previous Work and BIRCH Data clustering has been studied in the Machine Learning [2, 9, 10, 12, 21], Statistics [4, 5, 22, 23] and Database [6, 7, 24] communities with different methods and different emphases. 2.1. Probability Based Clustering Previous data clustering work in Machine Learning is usually referred to as unsupervised conceptual learning [2, 9, 12, 21] They concentrate on ....

....in the Machine Learning [2, 9, 10, 12, 21] Statistics [4, 5, 22, 23] and Database [6, 7, 24] communities with different methods and different emphases. 2.1. Probability Based Clustering Previous data clustering work in Machine Learning is usually referred to as unsupervised conceptual learning [2, 9, 12, 21]. They concentrate on incremental approaches that accept instances one at a time, and do not extensively reprocess previously encountered instances while incorporating a new concept. Concept (or cluster) formation is accomplished by top down sorting , with each new instance directed through a ....

[Article contains additional citation context not shown here]

John H. Gennari, Pat Langley, and Douglas Fisher, Models of Incremental Concept Formation, Artificial Intelligence, vol. 40, 1989, 11-61.

....m as bien de tipo extensional. El problema de la atenci on selectiva o selecci on de atributos ha sido abordado con anterioridad como un m etodo para mejorar la eficacia de sistemas de clasificaci on [18] o como un requisito adecuado para modelar sistemas cercanos al comportamiento humano [19]. Sin embargo, nunca se ha planteado como un mecanismo que permite modificar las descripciones de los conceptos que se manejan durante un proceso de aprendizaje. Para ilustrar esta perspectiva, se puede considerar un prototipo como una especie de descripci on continua. Esta descripci on viene dada ....

J. H. Gennari, P. Langley, and D. Fisher, "Models of incremental concept formation", Artificial Intelligence, vol. , pp. 11--61, 1989.

.... The most common means of supporting base filtering in CBR is to organise the case base as a decision tree that will support rough remindings without the need for exhaustive comparisons. Two approaches were considered for Cogger: Information theoretic: Using Gennari s Classit algorithm [8]. This method takes a case represented as a vector of numeric attributes and can incrementally locate the case in a classification hierarchy. In case retrieval this classification can produce the candidate set required from base filtering. The main advantage of Classit is that it is incremental, ....

....abstract profiles are more expensive to set up. From the perspective of the plagiarism detection task, the main novelty of this system is that it operates without doing exhaustive comparisons. Cogger performs similarity assessment as a two stage process; the first stage uses the Classit algorithm [8] to produce the candidate set of cases. The second stage performs expensive comparisons of the program structures to produce a metric of similarity (see Appendix I for details) The main conclusions from this exercise are as follows: Effective profiling is crucial: no amount of cleverness in ....

Gennari J. H., Langley P., Fisher D., "Models of incremental concept formation", Artificial Intelligence, pp11-61, Vol. 40, September, 1989.

....The source networks had 26 patterns, and were trained for 100 epochs. 4.1. 5 Heart disease diagnosis for Switzerland patients (Heart VAS, Heart HS) Using a 14 attribute set of diagnosis information, networks were trained for a heart disease diagnosis problem [ Schaffer, 1992, Detrano et al. 1989, Gennari et al. 1989, Aha et al. 1991 ] Input features included the following attributes: age, sex, chest pain type, resting blood pressure, serum cholesterol, fasting blood sugar, resting electrocardiographic results, maximum heart rate achieved (on a stress test) an exercise induced angina indicator, ST ....

J. H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--61, 1989.

....speaker. 24 hidden units were used, with j = 1:1; ff = 0:75. The target task had 20 patterns. 4. Heart disease diagnosis for Switzerland patients, transfer to Hungary: Using a 14attribute set of diagnosis information, we trained networks on a heart disease diagnosis problem [ Detrano et al. 1989, Gennari et al. 1989 ] We used networks trained on two different source problems, represented by data from a hospital in Hungary. The target problem was to train a network to perform the diagnosis task on Swiss patients. 7 hidden units were used, with j = 1:1; ff = 0:55. The target task had 1230 patterns. 5. Heart ....

J. H Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--61, 1989.

....(probability density functions) for continuous valued features are a much more difficult task. Therefore, the open problem in conceptual clustering is the problem of handling continuous valued and mixed continuous nominal valued data. The existing systems, such as COBWEB 3[19] AUTOCLASS, CLASSIT[12] have been developed to address this problem, but as the experimental results indicate, they produce much better results for nominal valued data sets than they do for mixed and continuous valued data. In this report, three methods developed to integrate schemes for handling combinations of numeric ....

J.H. Gennari, P. Langely, and D. Fisher. "Models of incremental concept formation," in Artificial Intelligence, 40:11-61, 1989.

....it is merely illustrative. Work Not Covered: Work not covered includes automatic construction of hierarchical structures using data in which the categories of objects are not known (unsupervised learning) present in fields such as cluster analysis [127] and machine learning (e.g. [141, 165]) A body of work using decision trees as a representational paradigm, existing in fields such as programming languages and analysis of algorithms, is not included. Work on decision trees constructed by hand (prevalent in the medical domain) is also not considered. 9 2.1.1 Outline of the ....

G. H. Gennari, Pat Langley, and Douglas Fisher. Models of incremental concept formation. Artificial Intelligence, 40(1--3):11--62, September 1989.

....because the probability that a continuous attribute will take on a particular number is zero. For such attributes, probabilities are estimated by assuming that values conform to a normal distribution with a particular mean and standard deviation. Thus, for a domain with continuous attributes, Gennari, Langley, and Fisher (1989) define category utility as 1 K K X k P (C k ) I X i 1 oe ik Gamma I X i 1 oe ip # ; 2) where P (C k ) is the probability of class C k , K is the number of categories, oe ik is the standard deviation for an attribute i in class C k , and oe ip is the standard deviation for ....

....concept hierarchy to weight attributes and bias retrieval. In summary, Cobweb s use of a hierarchical memory gives it the ability to represent complex target concepts, at least in principle. But whether its approach works in practice is an empirical issue, to which we now turn. 2. As discussed by Gennari et al. 1989), the value of 1 oe is undefined for any concept based on a single instance, so an acuity parameter is needed. Acuity corresponds to the notion of just noticeable difference in psychology. Learning Probabilistic Concept Hierarchies Page 6 3. Empirical studies of Cobweb Cobweb appeared on the ....

Gennari, J. H., Langley, P., & Fisher, D. H. (1989). Models of incremental concept formation. Artificial Intelligence, 40 , 11--61.

....regardless of the form of structure in the data Of course, if such a tree does not reasonably capture structure in data, then we might expect this to be reflected in error rate and or post validation simplicity. Nonetheless, there are probably better measures of cost available. In particular, Gennari (1989) observed Table 8: Cost characteristics of unoptimized and optimized clustering before and after validation. Average and standard deviations over 20 trials. Characteristics that are ed are computed from the mean values of Leaves and EPL. Unoptimized Optimized Unvalidated Validated ....

....equality, thus making both perspectives more explicit to an analyst. We have also assumed that variables are nominally valued. There have been numerous adaptations of the basic PU function, other functions, and discretization strategies to accommodate numeric variables (Michalski Stepp, 1983a,b; Gennari et al., 1989; Reich Fenves, 1991; Cheeseman, et al., 1988; Biswas et al., 1994) The basic sorting procedure and the iterative optimization techniques can be used with data described in whole or part by numerically valued variables regardless of which approach one takes. The identification of numeric variable ....

Gennari, J., Langley, P., & Fisher, D. (1989). Models of incremental concept formation.

No context found.

GENNARY J., LANGLEY P., FISHER D. 1989. Model of Incremental Concept Formation. Artificial Intelligence Journal, Volume 40, 11-61.

No context found.

Gennari, J. H.; Langley, P.; and Fisher, D. 1989. Models of incremental concept formation. Artificial Intelligence 40:11--61.

No context found.

J.H. Gennari, P. Langley, and D. Fisher, "Models of Incremental Concept Formation," Artificial Intelligence 40 (1989) 11-61.

No context found.

Gennari, J. H., Langley, P., Fisher, D. (1989). Models of incremental concept formation. Artificial Intelligence 40-1(3), pp.11-61.

No context found.

Gennari, J.H., Langley, P., Fisher, D. (1998). Models of incremental concept formation. Artificial Intelligence, 40(1-3):11-62.

No context found.

J.H. Gennari, P. Langley and D. Fisher. "Model of Incremental Concept Formation". Artificial Intelligence 40 (1-3), pp. 11-61, 1989.

No context found.

Gennari JH, Langley P, Fisher D. Models of incremental concept formation. Artificial Intelligence, 1989; 40:11-61.

No context found.

G. H. Gennari, Pat Langley, and Douglas Fisher. Models of incremental concept formation. Arti#cial Intelligence, 40#1#3#:11#62, September 1989.

No context found.

Gennari, J., Langley, P. and Fisher, D. (1989). Models of incremental concept formation. Artificial Intelligence Journal, vol. 40, pp. 11-61.

No context found.

GENNARI J., LANGLEY P., FISHER D. 1989. Model of Incremental Concept Formation.

No context found.

J. H. Gennari, P. Langley, and D. Fisher. Models of incremental concept formation. Artificial Intelligence, 40:11--61, 1988.

No context found.

Freeman. 13 Gennari, J. H., Langley, P. and Fisher, D. (1989). Models of Incremental Concept Formation. Artificial Intelligence 40: pp. 11-61.

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