Compact Rule Sets Research Articles

A fast scalable classifier tightly integrated with RDBMS

In this paper, we report our success in building efficient scalable classifiers by exploring the capabilities of modern relational database management systems (RDBMS). In addition to high classification accuracy, the unique features of the approach include its high training speed, linear scalability, and simplicity in implementation. More importantly, the major computation required in the approach can be implemented using standard functions provided by the modern relational DBMS. Besides, with the effective rule pruning strategy, the algorithm proposed in this paper can produce a compact set of classification rules. The results of experiments conducted for performance evaluation and analysis are presented.

μARTMAP: use of mutual information for category reduction in Fuzzy ARTMAP

A new architecture called muARTMAP is proposed to impact a category proliferation problem present in Fuzzy ARTMAP. Under a probabilistic setting, it seeks a partition of the input space that optimizes the mutual information with the output space, but allowing some training error, thus avoiding overfitting. It implements an inter-ART reset mechanism that permits handling exceptions correctly, thus using few categories, especially in high dimensionality problems. It compares favorably to Fuzzy ARTMAP and Boosted ARTMAP in several synthetic benchmarks, being more robust to noise than Fuzzy ARTMAP and degrading less as dimensionality increases. Evaluated on a real-world task, the recognition of handwritten characters, it performs comparably to Fuzzy ARTMAP, while generating a much more compact rule set.

Hierarchical fuzzy partition for pattern classification with fuzzy if-then rules

Fuzzy classification with fuzzy if-then rules suffers from the problem of pattern space partitioning, to reach a compact set of rules that provides high classification power. In this paper, we propose a hierarchical fuzzy partition method based on 2 N -tree decomposition. This decomposition is controlled by the grade of certainty of the generated rules for each fuzzy subspace and the deeper hierarchical level allowed. The performance of the proposed method is examined by the computer simulation using Iris data of Fisher. The proposed method is also compared with other fuzzy classification methods.

RIEVL: recursive induction learning in hand gesture recognition

Presents a recursive inductive learning scheme that is able to acquire hand pose models in the form of disjunctive normal form expressions involving multivalued features. Based on an extended variable-valued logic, our rule-based induction system is able to abstract compact rule sets from any set of feature vectors describing a set of classifications. The rule bases which satisfy the completeness and consistency conditions are induced and refined through five heuristic strategies. A recursive induction learning scheme in the RIEVL algorithm is designed to escape local minima in the solution space. A performance comparison of RIEVL with other inductive algorithms, ID3, NewID, C4.5, CN2, and HCV, is given in the paper. In the experiments with hand gestures, the system produced the disjunctive normal form descriptions of each pose and identified the different hand poses based on the classification rules obtained by the RIEVL algorithm. RIEVL classified 94.4 percent of the gesture images in our testing set correctly, outperforming all other inductive algorithms.

Extracting rules from neural networks by pruning and hidden-unit splitting.

An algorithm for extracting rules from a standard three-layer feedforward neural network is proposed. The trained network is first pruned not only to remove redundant connections in the network but, more important, to detect the relevant inputs. The algorithm generates rules from the pruned network by considering only a small number of activation values at the hidden units. If the number of inputs connected to a hidden unit is sufficiently small, then rules that describe how each of its activation values is obtained can be readily generated. Otherwise the hidden unit will be split and treated as output units, with each output unit corresponding to an activation value. A hidden layer is inserted and a new subnetwork is formed, trained, and pruned. This process is repeated until every hidden unit in the network has a relatively small number of input units connected to it. Examples on how the proposed algorithm works are shown using real-world data arising from molecular biology and signal processing. Our results show that for these complex problems, the algorithm can extract reasonably compact rule sets that have high predictive accuracy rates.

The analysis of relationships in databases for rule derivation

Owing to the rapid growth in the sizes of databases, potentially useful information may be embeded in a large amount of data. Knowledge discovery is the search for semantic relationships which exist in large databases. One of the main problems for knowledge discovery is that the number of possible relationships can be very large, thus searching for interesting relationships and reducing the search complexity are important. The relationships can be represented as rules which can be used in efficient query processing. We present a technique to analyze relationships among attribute values and to derive compact rule set. We also propose a mechanism and some heuristics to reduce the search complexity for the rule derivation process. An evaluation model is presented to evaluate the quality of the derived rules. Moreover, in real world, databases may contain uncertain data. We also propose a technique to analyze the relationships among uncertain data and derive probabilistic rules.

Direct induction of compact rule-based classifiers for resource mapping

Abstract Rule-based classifiers are used regularly with geographical information systems to map categorical attributes on the basis of a set of numeric or unordered categorical attributes. Although a variety of methods exist for inducing rule-based classifiers from training data, these tend to produce large numbers of rules when the data has noise. This paper describes a method for inducing compact rule-sets whose classification accuracy can, at least in some domains, compare favourably with that achieved by larger less succinct rule-sets produced by alternative methods. One rule is induced for each output class. The condition list for this rule represents a box in n-dimensional attribute space, formed by intersecting conditions which exclude other classes. Despite this simplicity, the classifier performed well in the test application prediction of soil classes in the Port Hills, New Zealand, on the basis of regolith type and topographic attributes obtained from a digital terrain model.

Rules for multidimensional multirate structures

Identifies a comprehensive set of compact rules and efficient algorithms for simplifying and rearranging structures common in multidimensional multirate signal processing. The authors extend the 1D rules reported by Crochiere and Rabiner (1983), especially the many equivalent forms of cascades of upsamplers and downsamplers. They also include rules reported by other authors for completeness. The extension to mD is based primarily on the Smith form decomposition of resampling (nonsingular integer square) matrices. The Smith form converts non-separable multidimensional operations into separable ones by means a shuffling of input samples and a reshuffling of the separable operations. Based on the Smith form, the authors have developed algorithms for 1) computing coset vectors 2) finding greatest common sublattices 3) simplifying cascades of up/downsampling operations. The algorithms and rules are put together in a form that can be implemented efficiently in a symbolic algebra package. The authors have encoded the knowledge in the commercially available Mathematica environment. >

Optimized rule induction

Swap-1, a state-of-the-art system for learning decision rules from data, is presented. The method embodied in Swap-1 generates reduced-complexity solutions by inducing compact solutions in larger dimensions where many rules might be needed to make accurate predictions. For many applications, such systems can automatically construct relatively compact rule sets with highly predictive performance. >

Jahn-Teller enhanced fundamental bands in the infrared spectra of degenerate electronic states

Scharf and Miller have shown recently that the symmetry selection rule of Child and Longuet-Higgins for infrared vibrational transitions in degenerate electronic states is incomplete, and that vibrational transitions that do not satisfy their selection rule can be exceptionally strong because the intensity is enhanced by the Jahn-Teller effect. The present paper provides a compact set of symmetry rules for the Jahn-Teller-enhanced vibrations, and considers the general rotational structure for the three types of Jahn-Teller-enhanced fundamentals, namely symmetric-top perpendicular bands, symmetric-top parallel bands, and spherical-top bands. The detailed equations of Child and Longuet-Higgins for the symmetric-top perpendicular case are valid, including the Jahn-Teller enhancement, but their generalization to other cases does not take account of the distinctive symmetry properties of the Jahn-Teller enhanced contribution. In the present calculations the use of an effective dipole moment operator to allow for the mixing of states by the Jahn-Teller interaction shows that the symmetry of the enhanced transitions is associated with the antisymmetric square of the electronic symmetry species, whereas the normal contributions considered by Child and Longuet-Higgins are associated with the symmetric square.