Tomofumi Nakano

A Mining Algorithm Using Property Items Extracted from Sampled Examples

This paper proposes a mining algorithm for relational frequent patterns based on a bottom-up property extraction from examples. The extracted properties, called property items, are used to construct patterns by a level-wise way like Apriori. The property items are assumed to have a special form, which is defined in terms of mode declaration of predicates. The algorithm produces frequent itemsets as patterns without duplication in the sense of logical equivalence. It is implemented as a system called Mapix and is evaluated with four different datasets with comparison to Warmr . Mapix had large advantage in runtime.

Inducing Shogi Heuristics Using Inductive Logic Programming

This paper reports the results of an inductive logic programming (ILP) application to solve shogi or Japanese chess mating problems, which are puzzles using shogi rules. The problems can be solved by heuristic search of AND-OR trees. We propose a method of using the ILP technique to generate heuristic functions, which are automatically tuned according to the confidence of the knowledge induced by ILP. Experiments show that the method prunes search space compared with a naive search.

Relational pattern mining based on equivalent classes of properties extracted from samples

This paper extends the bottom-up relational miner Mapix[9]. It takes a relational database consists of multiple relational tables including a target relation, and enumerates patterns with which a large part of instances in the target relation match. The patterns are given as logical formulae. Although a well-known system Warmr generates and tests possible patterns, it has limitation in its efficiency. Mapix took a bottom-up approach and gained efficiency at the cost of variety of patterns. It searches and propositionalizes features appeared in instances. Patterns produced is only simple combinations of attributed. The proposed algorithm EquivPix (an equivalent-class-based miner using property items extracted from examples) keeps the merits of bottom-up approach, i.e. time-efficiency and prohibition of duplicated patterns, and it widens pattern variation. EquivPix introduces equivalent classes on properties extracted and also two combination operators of them.

Relational association mining based on structural analysis of saturation clauses

Restricting the form of rules is an important issue of relational association rule mining. The proposing method PIX extracts properties from given examples and to use them to form rules. An property of an instance consists of an addressing part which specifies objects related to the instance and description part which says something among the objects. Extracted properties are used like as an item in market basket database and an APRIORI-like algorithm calculates frequent item sets. The paper describes also an experiment in a sample application.

e-Learning materials development based on abstract analysis using web tools

This study includes an original corpus of engineering journals and is part of the series of E-Learning & English for Specific Purposes (ESP) researches . Purposes (ESP) researches that includes an original corpus of engineering journals. In this paper the results of a corpus study will be presented, and a sample of the ESP e-learning materials being developed for graduate students in engineering will be shown. Abstracts were chosen for the corpus this time because students are likely to read many for their research, and eventually to have to produce their own. We prepare the 40,000-word corpus that consists of 263 abstracts from mechanical and electrical engineering journals. The corpus is analyzed using Wmatrix, which gives part-of-speech tags and semantic tags, and compares the results with those of the BNC written corpus sampler. Some special features found in the analysis are frequencies in semantic tags, part-of-speech tags, difference in the use of verbal forms and multi-words. As an application of the important features, we are developing web-based materials which include the original abstracts with target items hyper-linked to various pages containing exercises, concordances, grammar explanations, a bilingual dictionary, etc.

Solving Selection Problems Using Preference Relation Based on Bayesian Learning

This paper defines a selection problem which selects an appropriate object from a set that is specified by parameters. We discuss inductive learning of selection problems and give a method combining inductive logic programming (ILP) and Bayesian learning. It induces a binary relation comparing likelihood of objects being selected. Our methods estimate probability of each choice by evaluating variance of an induced relation from an ideal binary relation. Bayesian learning combines a prior probability of objects and the estimated probability. By making several assumptions on probability estimation, we give several methods. The methods are applied to Part-of-Speech tagging.

Validation of 3D Convolutional Neural Networks for Customer Satisfaction Estimation Using Videos

Companies and public entities administer customer satisfaction surveys to identify business problems. However, devising and analyzing questionnaires is burdensome for investigators, and answering questionnaires imposes a burden on customers. In addition, the response rate is frequently low. Here, to address these problems, we attempt to estimate customer satisfaction using sensing technology. We hypothesize that satisfaction can be discerned through facial expressions and body movements. To validate this hypothesis, we applied three-dimensional convolution neural networks.

Control of Hypothesis Space Using Meta-knowledge in Inductive Learning

Inductive logic programming (ILP) is effective for classification learning because it constructs hypotheses combining background knowledge. On the other hand it makes the cost of search for hypothesis large. This paper proposes a method to prune hypothesis using a kind of semantic knowledge. When an ILP system uses a top-down search, after it visits a clause (rule) it explore another clause by adding a condition. The added condition may be redundant with other conditions in the clause or the condition may causes the body of clause unsatisfied. We study to represent and use to treat the redundancy and unsatisfactory of conditions as meta-knowledge of predicates. In this paper we give a formalism of meta-knowledge and show to use it with an ILP algorithm. We also study a method to generate meta-knowledge automatically. The method generates meta-knowledge which controls redundancy and contradiction with respect to predicates by testing properties extensionally.

Solving selection problems combining relational learning and Bayesian learning

This paper defines a selection problem where an appropriate object is selected from a set that is specified by parameters. We discuss inductive learning of selection problems and proposed a method combining inductive logic programming (ILP) and Bayesian learning. Our methods estimate probability of each choice by evaluating likelihood of an induced relation from an ideal binary relation. Bayesian learning combines a prior probability of objects and the estimated probability. By making several assumptions on probability estimation, we give several methods. The methods are applied to Part-of-Speech tagging.

Organising documents based on standard-example split test

A purpose of text-mining is to summarise a large collection of documents. This paper proposes a new method to view a summary of large document set. It consists of two techniques, one of which constructs classification trees using a split test called the standard-example (standard-document) split test, and the other is a method to display features in each class of documents classified in the trees. The standard-example split test is a test which divides examples by their distance (or similarity) from a standard-example which is selected by a criterion. This is the first method which applies this test to text mining. The display method exhibits representative words of document classes which emphasise their feature.