This study uses an interview method to investigate how scientists use analogies in actual remarkable scientific discoveries. The observed cases of analogy are classified into six types according to the two criteria of similarity and transfer, only four of which types actually include the observed cases. Each typical case of three of the four types is shown and explained in relation to the availability of the goal of the analogy and to other forms of reasoning. One of the three cases is especially considered a case of analogy with the reformation of knowledge based on thematic abstraction. whose dynamic process has not fully been investigated.

This paper describes applications of text data mining to interactive document browsing and web mining. We develop a fast text data mining method based on optimal pattern discovery, and make experiments on large collections of documents and Web pages to evaluate the proposed method.

We developed a Progol/Database integrated system called DB-Amp to realize data mining lightly against structured data. A new data model RER(Refined Entity Relationship model)is introduced to define a target concept to be learned as well as background knowledge. A data conversion system DAISY (Data Arrangement Interface SYstem)is developed to automate the preparation of input data to Progol as far as possible. We concentrate on the following issues: (1)given a database and an inductive inference problem, how to design target concept representation to be learned by the ILP system Progol, (2)how to identify necessary background knowledge(BK)and how to define predicates constituting the BK, and (3)how to restrict the Progol search space to a finite set. We applied DB-Amp to build a rule base for an automatic e-mail classification system in real business environment.

Since the late 1980's, studies on Knowledge Discovery in Databases(KDD)has been paid much attentions. Briefly speaking, KDD processes can be divided into the four processes ((1) data selection, (2)pre-processing(or data cleaning), (3)data mining, and (4)interpretation and evaluation). Among these, the pre-processing is regarded as an most important process to prevent KDD processes from extracting meaningless rules. Generalization of databases used in an attribute-oriented induction can be considered as one of useful methods of the pre-processing. An attribute-oriented induction is a useful data mining method that generalizes databases under an appropriate abstraction hierarchy to find meaningful konwledge. The hierarchy is well designed so as to exclude meaningless rules from a particular point of view. However, there may exist several ways of generalizing databases according to user's intention. It is therefore important to provide a multi-layered abstraction hierarchy under which several generalizations are possible and are well controlled. In fact, too-general or too-specific databases are inappropriate for mining algorithms to extract significant rules. From this viewpoint, we propose a generalization method based on an information theoretical measure to select an appropriate abstraction hierarchy. The hierarchy can be considered as a layered abstraction, which is defined as a grouping of attribute values at concrete level. Futhermore, we consider a data selection method to extract meaningful rules. The method controls weight values(called votes)to extract a subset of tuples in the original database. The subset should be selected so that it forms a relatively meaningful mass of tuples. Then the same discovery method is applied for the subset to produce a hidden rule for the subset, not for the whole database. Finally, we present a system, ITA(Information Theoretical Abstraction), based on the generalization and selection methods, and tested it for a census database to show the effectiveness and the validity of ITA.

This paper presents an algorithm for discovering pairs of an exception rule and a common sense rule under a prespecified schedule for updating threshold-values. An exception rule, which represents a regularity of exceptions to a common sense rule, often exhibits interestingness. Discovery of pairs of an exception rule and a common sense rule, which can be regarded as hypothesis-driven, has been successful in various domains. In this method, however, both the number of discovered rule-pairs and time needed for discovery depend on the values of thresholds, and an appropriate choice of them requires experetise on the data set and on the discovery algorithm. In order to circumvent this problem, we propose two scheduling policies for updating values of these thresholds based on a novel data structure. The data structure consists of multiple balanced search-trees, and efficiently manages discovered patterns with multiple indices. One of the policies represents a full specification of updating by the user, and the other iteratively improves a threshold-value by deleting the worst pattern with respect to its corresponding index. Preliminary results on four real-world data sets are highly promising: our algorithm settled values of thresholds appropriately, and discovered interesting exception-rules from all these data sets.

We propose General-purposive decision tree generation algorithm(MINI based TREE, named MINI-TREE), which guarantees to obtain quasi-optimum solution with less computational complexity even in worst condition to develop decision tree. In inductive learning, decision tree is widely known as one of the representative knowledge representation scheme. Because of effectiveness, decision tree is used in various application fields. ID3 and C4.5 are famous and typical top-down algorithms, which create the decision tree from the root node by evaluating the amount of gain information of each attribute and deciding which attribute is fitted in the active node. However, they can maintain its performance, only when each attribute is assumed to be semantically independent in attributes-selection scheme from gain information. Therefore, the performance of them remarkably lowers if there exists strong correlation between some attributes in a set of examples. MINITREE is a top-down algorithm that decides to select the attributes by the selection-criterion using the logical formula minimization process of logic minimization algorithm named MINI. By applying logic minimization algorithm, the selection of the attributes as an importance becomes possible in order to develop the compact decision tree in each attributes selection step, even if it is the attribute inferior in respect of gain information. Especially, MINITREE is shown to be much useful to mass data including strongly correlative attributes.

KeyGraph, an automatic document indexing method for extracting keywords expressing the assertions of a document, i. e. assertions supported by the outlines based on the basic concepts of the document, is applied to detecting risky active faults from earthquake history data. Here a history data is regarded as a document to be indexed, and active faults stressed strongly i. e. with near-future earthquake risks are obtained as keywords asserted in the document. This paper presents this method and its seismologic semantics. The semantics shows that KeyGraph is a model of earthquake occurrences, which considers less details of local land crust activities than in seismology, but more of global interactions among active faults. Experimentally, faults with near-future earthquake risks were obtained with high accuracies, and the shifts of risky areas after big earthquakes datected by KeyGraph corresponded with realistic tectonics.

For automatic extraction of essential information and discovery from massive time series, it is necessary to develop a method which is flexible enough to handle actual phenomena in real world. That can be achieved by the use of state space model, and it provides us with a unified and computationally efficient filtering method and treating missing observations. As an example of successful applications of the method, analysis of groundwater level data is shown. It is shown that various discoveries are obtained from massive and noisy time series.

A novel and generic theory is formulated to characterize the structure of a scientific law/model equation. Based on the theory, an efficient algorithm is developed to discover scientific law/model equations governing an objective process under experimental environments, and the algorithm is implemented to the "Smart Discovery System(SDS)"program. SDS derives the quantitative equations reflecting the scientific first principles underlying the objective process. The power of the proposed approach comes from the use of "scale-type constraints" to limit the mathematically admissible relations among the measurement quantities representing the states of the objective process. These constraints well specify the admissible formulae of the scientific law/model equations, and provide a measure to efficiently reduce the search space of the equation formulae. In this paper, the theoretical foundation to discover the scientific law/model equations and the algorithm of SDS are presented, and its efficiency and practicality are demonstrated and discussed with complex working examples. Since the conventional equation discovery systems could not sufficiently guarantee the mathematical admissibility of the discovered equations, this work is expected to open up a new research field on the scientific equation discovery.

One of the important steps in knowledge discovery is to access the desired information from the large mount of data stored on the WWW. At present, such information can be accessed by a browser itself or by using a keyword search function. However, browsing is a time consuming task where a user must access individual hypertexts one by one. Furthermore, a keyword search is realized in a very limited way ; it does not support database-like queries. In this paper, we provide a smooth, interactive information retrieval framework in which WWW information can be viewed in a global sense and can be handled using a database-like query. Global view displays the link structure of WWW in a hyperbolic tree style where many texts can be accessed easily by changing focus. The query function is useful to filter irrelevant texts and reduce the size of the link structure. These functions are integrated within our browser, allowing us to discover desired texts from large hypertexts incrementally. Typical three web sites were selected to show the performance of the proposed method with improved accuracy and efficiency in information access to the web sites.

Systematic understanding of biological phenomena is necessary for its modeling. However, traditional analyses with differential equations often ended in an imitation of observed laboratory results, and too much effort was spent in reproducing the known quantitative data. What is more important is a strategy to computationally reconstruct the systematic aspect of biology. The paper introduces a strategy to represent metabolism using a graph. Its key idea is simple: to model metabolism only with enzymatic reactions and with chemical structures at the atomic level. The strategy was realized as AMR system, and was tested for a simple tracer experiment around TCA cycle.

Human can easily interact with computer systems, such as query-answer systems, machine translation systems and so on, which equip the natural language interface. In these systems, the interface should understand various user's subjective requests in order to prevent him(her)from feeling a kind of barriers for a computer. In Japanese, we find a variety of expressions to express the subjectivity in daily. Temporal adverbs which are for denoting temporal concepts are one of subjective expressions. In this paper, we argue the semantic representation of temporal adverbs, especially, MOU(already)and MADA(still)which are typical words for expressing a speaker's subjectivity. We introduce the state representation in which the combination of several temporal concepts corresponds to the temporal knowledge expression of sentences. In our method, we use the speaker's subjective time and the moment of event. The moment of event represents the event which is described in a sentence. The speaker's subjective time represents the next state or the previous state of the moment of event. We found that the easiness of inferring the speaker's subjectivity which is from the nature of predicate in a sentence determines whether a sentence is natural or not. As an application, we show the query-answer system which uses knowledge representation of our time expression. As a result, we found that our method is familiar with the Hornstein's expression for expressing the practical temporal knowledge and useful for human-familiar natural language interface.

With the motive of establishing a contradiction-free abductive proof procedure for extended logic programs containing both explicit negations and negations as failure, we deal with a contradiction caused by complementary pairs in a model with respect to the explicit negation. Revising the usage of the exception in Kowalski and Sadri(1990), we present a way of eliminating the contradiction so that a transformation of a contradictory program can be performed to a contradiction-free program with revised exceptions. The semantics for the transformed program with revised exceptions can be defined in terms of alternating fixpoint as in Gelder(1993). Excluding incompatible complementary pairs for explicit negations, we propose an abduction framework based on 3-valued logic, where the integrity constraint is well-defined for the transformed program with revised exceptions. We find the close relationship between the satisfactory integrity constraint in the abduction framework and the semantic conditions for the transformed program. The contradiction-free abductive proof procedure is given in the propositional logic, for the satisfaction of the abduction framework constraints, where the soundness of the procedure is briefly presented.