Hiroki Nomiya

A shape-based similarity measure for time series data with ensemble learning

This paper introduces a shape-based similarity measure, called the angular metric for shape similarity (AMSS), for time series data. Unlike most similarity or dissimilarity measures, AMSS is based not on individual data points of a time series but on vectors equivalently representing it. AMSS treats a time series as a vector sequence to focus on the shape of the data and compares data shapes by employing a variant of cosine similarity. AMSS is, by design, expected to be robust to time and amplitude shifting and scaling, but sensitive to short-term oscillations. To deal with the potential drawback, ensemble learning is adopted, which integrates data smoothing when AMSS is used for classification. Evaluative experiments reveal distinct properties of AMSS and its effectiveness when applied in the ensemble framework as compared to existing measures.

Improvement of Boosting Algorithm by Modifying the Weighting Rule

AdaBoost is a method for improving the classification accuracy of a given learning algorithm by combining hypotheses created by the learning alogorithms. One of the drawbacks of AdaBoost is that it worsens its performance when training examples include noisy examples or exceptional examples, which are called hard examples. The phenomenon causes that AdaBoost assigns too high weights to hard examples. In this research, we introduce the thresholds into the weighting rule of AdaBoost in order to prevent weights from being assigned too high value. During learning process, we compare the upper bound of the classification error of our method with that of AdaBoost, and we set the thresholds such that the upper bound of our method can be superior to that of AdaBoost. Our method shows better performance than AdaBoost.

Introducing specialization and generalization to a graph-based data model

This paper proposes the schema graph for introducing specialization and generalization to a graph-based data model in order to systematize and reuse knowledge effectively. Systematizing and reusing knowledge are important functions of the knowledge-based human activity. The schema graph enables specialization and generalization relationships to be dynamically added, and removed. The methods of modifying these relationships are precisely presented. The schema graph enables us to systematize and reuse knowledge with keeping the structure flexible.

Impression Evaluation Method Considering the Vagueness of Kansei

This paper proposes a method for evaluating impressions using a space containing impression words. The impression of an object is specified by circling impression words that match the impression. The degree of matching of the impression is expressed by the darkness of the color used to paint the circled area. This proposed method poses few restrictions to users evaluating an impression. Conversely, the semantic differential (SD) method evaluates an impression in a predefined range to enable statistical processing. A prototype system is implemented to evaluate the proposed method, and experimentally compare it with conventional methods, such as the SD method. Finally, it is demonstrated that the proposed method can evaluate aspects that other methods cannot.

Obtaining Factors Describing Impression of Questions and Answers and Estimation of Their Scores from Feature Values of Statements

First, this paper evaluates the impression of questions and answers at Questions and Answers (Q & A) sites in order to avoid the problem of mismatch between the questioner and the respondent. Fifty impression words effective in evaluating impressive expression of statements are selected from a dictionary. An impressive evaluation experiment is then conducted for sixty questions and answers posted at Yahoo! Chiebukuro by using those impression words. Nine factors are obtained by applying factor analysis to the scores obtained through the experiment. Then factor scores of any other statements are tried to be estimated by using multiple regression analysis. This result, however, shows that the estimation accuracy is insufficient. To improve the estimation accuracy, the multiple regression analysis considering quadratic terms is applied. The result of the analysis shows that the estimation accuracy can be improved.

Effectiveness of the Analysis Method for the Impression Evaluation Method Considering the Vagueness of Kansei

In recent years, Kansei becomes important. However, the conventional methods are difficult to evaluate Kansei because Kansei is vague. An impression evaluation method considering the vagueness of Kansei has been proposed. This evaluation method makes a subject evaluate impression spatially. A method for analyzing the evaluation results has been proposed and the results of analysis have been shown. This analysis method shows average values and coefficients of variation of scores of the evaluation results spatially. However, only a few subjects joined the evaluation experiment. In this paper, an evaluation experiment is newly conducted, and more evaluation results are obtained. These results are analyzed, and characteristics of the impression of objects and the dispersion among subjects could easily be obtained. It is shown that this analysis method is useful for examining characteristics of impression of objects.

Generation Method of Concurrency Control Program by Using Genetic Programming

This paper proposes a generation system of concurrency control program by using genetic programming (GP). This system generates concurrency control program according to the features of transactions, which are collections of database operations. Functions and terminals of trees representing program in GP, and the fitness measure function used in GP are proposed. The functions and the terminals include those changing and testing variables attached to data items and transactions as well as those checking the kind of operation etc. These will bring us general concurrency control program, which is beyond the combination of the parts of traditional concurrency control program. As the granularity of the functions and the terminals is small, the sub-trees, which are used for the popular concurrency control protocol, and are prepared in advance, are used. The fitness measure function considers the goodness of concurrency control program. The experiments show that a concurrency control program using locks could be generated under the concurrent environment, while a concurrency control program better than the two-phase locking protocol could be generated under the not-so-concurrent environment.

Fast Subsequence Matching in Plasma Waveform Databases

This paper proposes a method of subsequence matching of plasma waveforms. The proposed method divides a waveform into fine-grained segments. The similar segments are grouped into a segment group. A multi-dimensional index is used for quick retrieval. Grouping segments could save the amount of the index. In the retrieval, a sequence of segments, which is called a section, is used as a unit in matching subsequences. Overlapping sections could overcome the shift errors of subsequences, and results in good retrieval correctness.

Multistrategical image classification for image data mining

For an efficient image data mining, accurately finding and retrieving various types of images are required. Therefore, there is a need for an image classification method which can be widely applicable to image data mining tasks. But traditional methods can be applied only limited domains. In this paper, we propose a flexible and accurate image classification method. Our method adopts a visual learning framework, which is an effective image classification framework based on machine learning. Currently, most of visual learning methods adopt monostrategy learning frameworks using a single learning algorithm. However, the real-world objects are too complex to be correctly recognized by a monostrategy method. Thus, utilizing a wide variety of features is essential to precisely discriminate them. In order to utilize various features, we propose multistrategical visual learning by integrating multiple visual learners. In our method, a visual learner is trained using the examples misclassified by the other visual learners. Therefore, all the visual learners can be collaboratively trained. This complementary learning framework leads to a more efficient classification.

Impressive Scene Detection from Lifelog Videos by Unsupervised Facial Expression Recognition

In order to retrieve impressive scenes from life log videos, we propose an emotional scene detection method based on facial expression recognition. Many researches about facial expression recognition focus on discriminating typical facial expressions such as happiness, sadness and surprise. But they are not suitable for life log videos because more complicated or subtle facial expressions are frequently observed in them. The proposed method tries to solve this problem by constructing a facial expression recognition model based on unsupervised learning. It discriminates facial expressions and detects emotional scenes by a clustering approach using facial features based on the positional relationships of several facial feature points. This approach is fully flexible to detect various facial expressions from life log videos because it does not need to predefine the facial expressions. The detection performance of the proposed method is evaluated in terms of detection accuracy and efficiency through the emotional scene detection experiments.