Showing papers on "Multiclass classification published in 1997"
TL;DR: The paper presents a novel approach to case-based classification based on a notion of similarity assessment that was developed for supporting flexible retrieval of relevant information and is compared to that of other machine learning algorithms.
Abstract: Classification involves associating instances with particular classes by maximizing intra-class similarities and minimizing inter-class similarities. Thus, the way similarity among instances is measured is crucial for the success of the system. In case-based reasoning, it is assumed that similar problems have similar solutions. The case-based approach to classification is founded on retrieving cases from the case base that are similar to a given problem, and associating the problem with the class containing the most similar cases. Similarity-based retrieval tools can advantageously be used in building flexible retrieval and classification systems. Case-based classification uses previously classified instances to label unknown instances with proper classes. Classification accuracy is affected by the retrieval process – the more relevant the instances used for classification, the greater the accuracy. The paper presents a novel approach to case-based classification. The algorithm is based on a notion of similarity assessment and was developed for supporting flexible retrieval of relevant information. Case similarity is assessed with respect to a given context that defines constraints for matching. Context relaxation and restriction is used for controlling the classification accuracy. The validity of the proposed approach is tested on real-world domains, and the system's performance, in terms of accuracy and scalability, is compared to that of other machine learning algorithms.
23 citations
Patent•
23 May 1997TL;DR: In this paper, the authors proposed a method and an arrangement for pattern recognition on the basis of statistics, where the association with each target class of the class set is estimated with a numerical value which is produced by cascaded use of polynomial classifiers.
Abstract: The invention relates to a method and an arrangement for pattern recognition on the basis of statistics. According to said method, for an object to be recognised on the basis of a complete set of two-class or multiclass classifiers the association with each target class of the class set is estimated with a numerical value which is produced by cascaded use of polynomial classifiers. According to the invention, on a learning sample in which all class patterns to be recognised are represented to a sufficiently significant extent there is a selection, from all the two-class or multiclass classifiers by way of their estimation vector spectrum, of those two-class or multiclass classifiers with estimations contributing the most to minimise a scalar quantity calculated over the estimation vector spectrum and having high separating relevance. The selected two-class or multiclass classifiers are subsequently used to form, via an expanded learning sample, estimation vectors from which expanded characteristic vectors are produced by polynomial linking. An evaluation classifier is formed on the basis of said characteristic vectors for estimating all target classes.
21 citations
TL;DR: The classification of predicates and languages is compared with the classification of arbitrary recursive functions and with their learnability and a formalization of multi-classification is presented and completely characterized in terms of standard classification.
Abstract: The present paper studies a particular collection of classification problems, i.e., the classification of recursive predicates and languages, for arriving at a deeper understanding of what classification really is. In particular, the classification of predicates and languages is compared with the classification of arbitrary recursive functions and with their learnability. The investigation undertaken is refined by introducing classification within a resource bound resulting in a new hierarchy. Furthermore, a formalization of multi-classification is presented and completely characterized in terms of standard classification. Additionally, consistent classification is introduced and compared with both resource bounded classification and standard classification. Finally, the classification of families of languages that have attracted attention in learning theory is studied, too.
14 citations
TL;DR: A new initialization method is proposed that minimizes the initial objective function and eliminates the phenomenon that weights decrease in the beginning of learning and can be used with any accelerated learning algorithm to enhance the learning speed.
Abstract: Initial learning process of the BP, which can influence the performance of learning in multiclass classification problems, is analyzed. Also, the weights decreasing phenomena in the initial stage of learning are investigated. On the basis of this analysis, a new initialization method is proposed. The proposed method minimizes the initial objective function. It eliminates the phenomenon that weights decrease in the beginning of learning. Several simulation results show that the proposed initialization method performs much better than the conventional random initialization method in the batch mode and slightly better in the pattern mode. Since it requires only a little additional computation, it is a strong alternative to the conventional random initialization. It is expected that the proposed initialization method can be used with any accelerated learning algorithm to enhance the learning speed.
8 citations
Patent•
27 Jul 1997
TL;DR: In this article, an evaluation function by weighting or the like obtained from the rate of correct answer for the past classification is made to be a database by each classification personnel and, at the same time, a typical document for characterizing the classification by each classifier is prepared in advance.
Abstract: PROBLEM TO BE SOLVED: To perform a more precise classification of an object document by using both classification results by a manual classification and an automatic classification. SOLUTION: An evaluation function by weighting or the like obtained from the rate of correct answer for the past classification is made to be a database by each classification personnel and, at the same time, a typical document for characterizing the classification by each classification is prepared in advance. Then, a scoring of a manual classification of each classification is executed from a classification result by the classification personnel (manual) and the evaluation function of the object document. Also, the degree of similarity of the object document and the typical document is calculated and the scoring of an automatic classification of each classification is performed by using this degree of similarity. The classification of the highest value obtained by summing up both of these scores by each classification is regarded as the final classification result. In this way, by fusing the manual classification and the automatic classification, a more precise classification result can be obtained.
6 citations
TL;DR: In this paper, two modifications of the full multinomial classification rule for the discrete feature discrimination problem are proposed, one derived from a likelihood ratio test and the other from the Bayesian predictive approach.
Abstract: In this paper two modifications of the full multinomial classification rule for the discrete feature discrimination problem are proposed. One is derived from a likelihood ratio test and the other from the Bayesian predictive approach. A simulation experiment was conducted to compare the performance of the new rules with two other existing rules for the same problem: the usual estimative or plug-in multinomial rule and the Dillon-Goldstein rule which is based on a distance principle. For equal group sample sizes all rules are equivalent, however for different sizes they behave differently. The simulation study has shown that the three alternative rules lead to similar results, in terms of expected actual error rates, which are better than those for the plug-in rule.
5 citations
12 Oct 1997
TL;DR: A new optimization method for multiclass feature extraction problems by assigning weights to each class in computing the global criterion function and adjusting the weights as new features are extracted by using the error prediction technique.
Abstract: In this paper, we propose a new optimization method for multiclass feature extraction problems by assigning weights to each class in computing the global criterion function and adjusting the weights as new features are extracted. Recently, it is shown that it is possible to predict the classification error within 1-2% margin from the Bhattacharyya distance. We use the error prediction technique to adjust the weights of each classes. Initially, we assign equal weights to each class. After the first feature is extracted, we calculate classification error of each class when the first feature is used and adjust the weights accordingly. We compute again the global criterion function with a new set of weights excluding the first feature and calculate the second feature from the revised criterion function, and so on. Preliminary experiments show improvement over the conventional methods.
4 citations