Maximum Entropy Model (MEM) [1] [4] estimates probability distribution functions, by which current state of knowledge is described in the context of prior data. Here we examine Generalized Iterative Scaling (GIS) [1] algorithm to determine optimum feature weights with feature selection during learning. Maximum Entropy principle [1] provides us with all the characteristics of the data given in advance and we could expect robust distribution against outlier. However it takes much time until convergence because the computation depends heavily on the number of classes. We introduce a novel approach random sampling of Monte Carlo method into GIS for improved computation.

Reinforcement Learning allows us to acquire knowledge without any training data. However, for learning it takes time. In this work, we propose a method to perform Reverse action by using Retrospective Kalman Filter that estimates the state one step before. We show an experience by a Hunter Prey problem. And discuss the usefulness of our proposed method.

Social networking services (SNS) are important sources of news for the world. By summarizing text on SNS, we will be able to grasp what topics in the news are going on. However, because texts keep growing in SNS, we need to process huge amounts of data very efficiently. Therefore, it is necessary to apply a fast and low-memory algorithm. In this investigation, we propose Successive KeyGraph as a method of keywords extraction. The extracted keywords work effectively for news summarization. In Successive KeyGraph, we update keygraph over time. KeyGraph extracts the author's claims and outputs the results immediately. Also, we hold co-occurrence count which we need to make a keygraph current by using Lossy Counting. We apply several preprocessing to manage unknown words and to improve morphological analysis.

In this investigation, we discuss a game of pursuit-evasion, or a hunter-prey problems using Q-learning framework. This has always been a popular research subject in the field of robotics where a hunter moves around in pursuit a prey. We involve Kalman filters to estimate the prey's status (location and velocity) and learn Q-values based on the estimated status. We evaluate our approach by convergence of Q-values and capturing steps.

For the purpose of text classification or information retrieval, we apply preprocessing to these texts such as stemming and stopwords removal. Almost all the techniques could be useful only to well-formed text information like textbooks and news articles, but is not true to social network services (SNS) or any other texts in internet world. In this investigation, we propose how to extract stopwords in context of social network services. To do that, first we discuss what stopwords mean, how different from conventional ones, and we propose statistical filters TFIG and TFCHI, to identify. We examine categorical estimation to extract characteristic values putting our attention on Kullback Leibler Divergence (KLD) over temporal sequences on SNS data. Moreover we apply several preprocessing to manage unknown words and to improve morphological analysis.

In this investigation, we discuss high order data structure (called tensor) for efficient information retrieval and show especially how well reduction techniques of dimensionality goes while preserving Euclid distance between information. High order data structure requires much amount of space. One of the effective approaches comes from dimensionality reduction such as Latent Semantic Indexing (LSI) and Random Projection (RP) which allows us to reduce complexity of time and space dramatically. The reduction techniques can be applied to high order data structure. Here we examine High Order Random Projection (HORP) which provides us with efficient information retrieval keeping feasible dimensionality reduction.

In this investigation, we propose how to accelerate Q-learning which is one of the most successful reinforcement learning methods using mirror images for hunting problems. Mirror images have symmetric differences of views, and they allow us to accelerate Q-learning dramatically. In this investigation, we show Q-learning goes 2 times faster (one mirror) or 3 times faster (2 mirrors) but a capturing ability decreased slightly even if intervention happens. Moreover, we prove that the new approach gets to the convergence if the one with no mirror does.

In this paper we propose how to accelerate Q-learning which is one of the most successful reinforcement learning methods using mirror images for hunting problems. Mirror images have symmetric differences of views, and they allow us to accelerate Q-learning dramatically. In the results we show Q-learning goes 2 times faster (one mirror) or 3 times faster (2 mirrors) but a few percents of capturing ability decreases. Moreover, we prove that the proposed approach converges if the one with no mirror does.

Hierarchical Hidden Markov Models (HHMMs) are sophisticated stochastic models that enable us to capture a hierarchical context characterization of sequence data. However, existing HHMM parameter estimation methods require large computations of time complexity O(TN^{2D}) at least for model inference, where D is the depth of the hierarchy, N is the number of states in each level, and T is the sequence length. In this paper, we propose a new inference method of HHMMs for which the time complexity is O(TN^{D+1}). A key idea of our algorithm is application of the forward-backward algorithm to ''state activation probabilities''. The notion of a state activation, which offers a simple formalization of the hierarchical transition behavior of HHMMs, enables us to conduct model inference efficiently. We present some experiments to demonstrate that our proposed method works more efficiently to estimate HHMM parameters than do some existing methods such as the flattening method and Gibbs sampling method.

In this work, we discuss how an intelligent agent learns in combinatorial auctions. It is well-known that it is NP-complete problem to find the optimal allocation to maximize revenue, because this is a typical form of Set Package Problem (SPP). We apply a framework of machine learning to combinatorial auctions, and discuss how to extract intelligence about bidding behavior. We show empirical convergence of knowledge within Q-learning framework. By this result, we target fully automated negotiation systems.

In this work we discuss how to extract domain specific synonyms automatically. It is not easy to extract these synonyms or related words without hand-coding. We identify pairs of sentences from two corpus in Japanese, estimate translation probabilities and extract synonyms using statistical machine translation techniques.

In this investigation, we propose a probabilistic approach for estimating the ages of blog authors in Japan by means of naive Bayesian classifier. We can learn context of characteristic words appeared in training data in terms of entropy. The key idea is that we extract feature words specific to authors' ages, and we estimate ages of the blog authors. We show the effectiveness of our approach by experimental results.

In this investigation, we propose an experimental approach for word segmentation in Japanese under domain-dependent situation. We apply Conditional Random Fields (CRF) to our issue. CRF learns several probabilistic parameters from training data with specific feature functions dependent on domains. Here we propose how to define domain specific feature functions.

Dynamic hash allows us to adjust the size of hash space dynamically according to the amount of data so that we obtain the nice time/space efficiency of the hash space. One of the proposed techniques is linear hash (LH). However, practically the technique doesn't always provide us with suitable results, especially in the case of collective (consecutive) inserts. In this investigation, we propose a new novel approach, called tree hash (TH), for the purpose of sophisticated dynamic hash processing. Here we assume distributed environment, i.e., many CPUs and huge amount of data storage connected through high speed network with each other. We show empirically the several excellent properties compared to LH*, a distributed LH

This paper proposes a new approach to observe, summarize and track events from a collection of news Web pages. Given a set of temporal Web pages, we obtain valid times-tamp from Web pages and detect events by means of clustering. Then we track events by using KeyGraph based on the clusters and abstract the clusters by using SuffixTree. We examine some experimental results and show the usefulness of our approach

Nowadays there have several applications on spatial information which manage high dimensional data. Whenever we examine nearest neighbor search in these applications by multi-dimensional indexing structure, very often we must access all pages if dimensionality exceeds about 10. This is known as curse of dimensionality that says any indexing structure is outperformed by simple linear search. In this investigation, for high dimensional data, we propose a sophisticated access mechanism based on extensible grid files with dimensionality reduction (DR) technique. We analyze error estimation caused by DR and recover the search space on original dimension. We examine nearest neighbor search and discuss some empirical results to show the usefulness of our approach.

In this investigation we propose a novel technique to classify melodies by using EM algorithm. Here we generate classifiers based on naive Bayesian technique during FM steps. And we discuss the experimental results with several features of melodies and show we can get efficient classifiers by our approach.

In document retrieval task, random projection (RP) is considered as a useful technique of dimension reduction. The task can be achieved very quickly yet the recalculation is not necessary to any changes. However, in lower dimension, random projection has instability by randomness in itself. In this investigation, we propose a new technique, called skewed projection (SP) for dimension reduction technique based term frequency distribution. We show that this technique can take advantages of local independency thus we can obtain efficient retrieval for documents which belong to specific application area.

We propose a new mechanism to give temporal order to a news article in a form of times-tamps. Here we learn temporal data in advance to extract ordering by means of incremental clustering and then we estimate most likely order to news text. In this work, we examine TDT2 corpus and we show how well our approach works by some experiments.

Here we discuss how to look for similar melody in music databases by giving monophonic melody in sheet. In this work, we utilize text expression (or sheet music) to describe music and introduce pitch spectrum of melodies. By this feature, we concisely distinguish music from tempo, transposition or other arbitrary expressions. We show the usefulness by experimental results.

In this work, we discuss how to interpret traditional Data Flow Diagram (DFD) by Unified Modeling Language (UML) for the purpose of integrating legacy systems with modern systems. To do that we introduce DFD meta model by using UML class diagrams and Object Constraint Language (OCL). We utilize the meta model to capture DFD semantics and establish close correspondence between DFD expressions and the expressions that can be derived by instantiating the meta framework. At the same time we describe procedures to generate UML descriptions from DFD based on the correspondence.

In this investigation, we discuss how to mine Temporal Class Schemes to model a collection of time series data. From the viewpoint of temporal data mining, this problem can be seen as discretizing time series data or aggregating them. Also this can be considered as screening (or noise filtering). From the viewpoint of temporal databases, the issue is how we represent the data and how we can obtain intensional aspects as temporal schemes. In other words, we discuss scheme discovery for temporal data. Given a collection of temporal objects along with time axis (called log), we examine the data and we introduce a notion of temporal frequent classes to describe them. As the main results of this investigation, we can show that there exists one and only one interval decomposition and the temporal classes related to them. Also we give experimental results that prove the feasibility to time series data.

Article Free Access Share on Extensible perfect hashing Authors: Takao Miura Dept. of Elect. and Elect. Engr., Hosei University, Kajinocho, Koganei, Tokyo, Japan Dept. of Elect. and Elect. Engr., Hosei University, Kajinocho, Koganei, Tokyo, JapanView Profile , Wataru Matsumoto Dept. of Elect. and Elect. Engr., Hosei University, Kajinocho, Koganei, Tokyo, Japan Dept. of Elect. and Elect. Engr., Hosei University, Kajinocho, Koganei, Tokyo, JapanView Profile , Isamu Shioya Dept. of Informatics, SANNO University, Kamikasuya, Isehara, Kanagawa, Japan Dept. of Informatics, SANNO University, Kamikasuya, Isehara, Kanagawa, JapanView Profile , Yukio Wada HITACHI Software Engineering Co. Ltd., Yokohama, Kanagawa, Japan HITACHI Software Engineering Co. Ltd., Yokohama, Kanagawa, JapanView Profile Authors Info & Claims CIKM '00: Proceedings of the ninth international conference on Information and knowledge managementNovember 2000 Pages 446–452https://doi.org/10.1145/354756.354852Published:06 November 2000Publication History 2citation298DownloadsMetricsTotal Citations2Total Downloads298Last 12 Months12Last 6 weeks1 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my AlertsNew Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteeReaderPDF

We propose a novel pruning method of decision trees based on domain knowledge, semantic hierarchies among classes, which is used to generate decision trees by relaxing the levels of hierarchies for both height and width of the trees. We develop the algorithm, and the effectiveness is examined by UCI Machine Learning Repository: On Car Evaluation and Nursery. We can generate the decision trees consisting of 11 and 13 rules, although C4.5 generates 182 and 572 rules, respectively.

In this investigation, we discuss optimization issues of meta-queries for databases. Since traditional techniques can't be applied to meta-queries, we propose a new technique of cost-based optimization without the evaluation of meta-objects. For this purpose, we introduce signature techniques and extend database schemes.

We propose a novel technique to learn concepts from a large set of objects by means of clustering techniques. Unlike conventional approaches, we utilize database scheme knowledge and specify what clusters mean by linking attribute-grouping to object clustering. We propose some sophisticated algorithms (and some extensions) and show how useful they are.

A scheme in databases can be considered as a knowledge description in database instances. To design schemes of databases is a knowledge acquisition technique. One of the deficiencies in database design is that there are no unique solutions. In addition, we can not avoid an evolution of the schemes because of the database paradigm. It is useful to employ AI methodologies, such as machine learning or knowledge discovery in database design. We assume ISA hierarchies for knowledge control, examine the data and present a knowledge acquisition method to discover or find the difference between type schemes, in a sense of approximation between a scheme knowledge and a knowledge obtained from database instances.

We propose a new method to mine a type scheme semi-automatically from an initial database scheme and the instances. Our data model assumes that one entity may have more than one type and classification (or type scheme). It might be appropriate when each entity is classified into at most k (least general) classes with respect to the ISA hierarchy, to keep database processing efficient. Our method differs from others in evolving ISA hierarchy by introducing a semantical metric. We propose a sophisticated algorithm to simplify, evolve and generate type schemes.

A data processing model called PIM is proposed. This model has its own database diagram mechanisms and a visual data manipulation language is defined on the diagrams. The visual language allows users to construct database queries utilizing iconic pictures over database diagrams. The algebraic capabilities of the language in query processing are described, and a formal definition is given. Related systems are reviewed, and the distinct features of the present system are highlighted.>