• Proceedings of the KSPS conference
• /
• 2005.04a
• /
• pp.87-90
• /
• 2005

Today's state-of the-art speech recognition systems typically use continuous distribution hidden Markov models with the mixtures of Gaussian distributions. To obtain higher recognition accuracy, the hidden Markov models typically require huge number of Gaussian distributions. Such speech recognition systems have problems that they require too much memory to run, and are too slow for large applications. Many approaches are proposed for the design of compact acoustic models. One of those models is subspace distribution clustering hidden Markov model. Subspace distribution clustering hidden Markov model can represent original full-space distributions as some combinations of a small number of subspace distribution codebooks. Therefore, how to make the codebook is an important issue in this approach. In this paper, we report some experimental results on various quantization methods to make more accurate models.

• Journal of Digital Convergence
• /
• v.20 no.2
• /
• pp.217-223
• /
• 2022

Subspace clustering for online data streams requires a large amount of memory resources as all subsets of data dimensions must be examined. In order to track the continuous change of clusters for a data stream in a finite memory space, in this paper, we propose a grid-based subspace clustering algorithm that effectively uses memory resources. Given an n-dimensional data stream, the distribution information of data items in data space is monitored by a grid-cell list. When the frequency of data items in the grid-cell list of the first level is high and it becomes a unit grid-cell, the grid-cell list of the next level is created as a child node in order to find clusters of all possible subspaces from the grid-cell. In this way, a maximum n-level grid-cell subspace tree is constructed, and a k-dimensional subspace cluster can be found at the k^th level of the subspace grid-cell tree. Through experiments, it was confirmed that the proposed method uses computing resources more efficiently by expanding only the dense space while maintaining the same accuracy as the existing method.

• Communications for Statistical Applications and Methods
• /
• v.27 no.4
• /
• pp.431-443
• /
• 2020

The K-means clustering algorithm has had successful application in sufficient dimension reduction. Unfortunately, the algorithm does have reproducibility and nestness, which will be discussed in this paper. These are clear deficits for the K-means clustering algorithm; however, the hierarchical clustering algorithm has both reproducibility and nestness, but intensive comparison between K-means and hierarchical clustering algorithm has not yet been done in a sufficient dimension reduction context. In this paper, we rigorously study the two clustering algorithms for two popular sufficient dimension reduction methodology of inverse mean and clustering mean methods throughout intensive numerical studies. Simulation studies and two real data examples confirm that the use of hierarchical clustering algorithm has a potential advantage over the K-means algorithm.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.34S no.12
• /
• pp.30-40
• /
• 1997

This paper presents a design method of fuzzy control systems by clustering the data in the subspace of the input-output produyct space. In the case of servo control, most input-outputdata are concentrated in thye steady-state region, and the the clustering will result in only steady-state fuzzy rules. To overcome this problem, we divide the input-output product space into some subspaces according to the state of input variables. The fuzzy control system designed by the subspace clustering showed good transient response and smaller steady-state error, which is comparable with the reference fuzzy system.

• ETRI Journal
• /
• v.37 no.2
• /
• pp.317-327
• /
• 2015

A reliable analysis of consumer preference from a large amount of purchase data acquired in real time and an accurate customer characterization technique are essential for successful direct marketing campaigns. In this study, an optimal segmentation of post office customers in Korea is performed using a subspace projection-based clustering method to generate an accurate customer characterization from a high-dimensional census dataset. Moreover, a traditional temporal mining method is extended to an algorithm using the MapReduce framework for a consumer preference analysis. The experimental results show that it is possible to use parallel mining through a MapReduce-based algorithm and that the execution time of the algorithm is faster than that of a traditional method.

• Journal of KIISE:Databases
• /
• v.32 no.3
• /
• pp.228-242
• /
• 2005

The projected clustering seeks to find clusters in different subspaces within a high dimensional dataset. We propose an algorithm to discover near optimal projected clusters without user specified parameters such as the number of output clusters and the average cardinality of subspaces of projected clusters. The objective function of the algorithm computes projected energy, quality, and the number of outliers in each process of clustering. In order to minimize the projected energy and to maximize the quality in clustering, we start to find best subspace of each cluster on the density of input points by comparing standard deviations of the full dimension. The weighting factor for each dimension of the subspace is used to get id of probable error in measuring projected distances. Our extensive experiments show that our algorithm discovers projected clusters accurately and it is scalable to large volume of data sets.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.17 no.4
• /
• pp.703-708
• /
• 2022

K-means clustering is a representative clustering technique. However, there is a limitation in not being able to integrate the performance evaluation scale and the method of determining the minimum number of clusters. In this paper, a method for numerically determining the minimum number of clusters is introduced. The explained variance is presented as an integrated measure. We propose that the k-means clustering method should be performed in the subspace of the PCA in order to simultaneously satisfy the minimum number of clusters and the threshold of the explained variance. It aims to present an explanation in principle why principal component analysis and k-means clustering are sequentially performed in pattern recognition and machine learning.

• The Korean Journal of Applied Statistics
• /
• v.35 no.5
• /
• pp.657-666
• /
• 2022

An informative predictor subspace is useful to estimate the central subspace, when conditions required in usual suffcient dimension reduction methods fail. Recently, for multivariate regression, Ko and Yoo (2022) newly defined a projective-resampling informative predictor subspace, instead of the informative predictor subspace, by the adopting projective-resampling method (Li et al. 2008). The new space is contained in the informative predictor subspace but contains the central subspace. In this paper, a method directly to estimate the informative predictor subspace is proposed, and it is compapred with the method by Ko and Yoo (2022) through theoretical aspects and numerical studies. The numerical studies confirm that the Ko-Yoo method is better in the estimation of the central subspace than the proposed method and is more efficient in sense that the former has less variation in the estimation.

• Journal of Korea Society of Industrial Information Systems
• /
• v.18 no.3
• /
• pp.35-46
• /
• 2013

In order to develop an accurate analytical model for domestic electricity demand forecasting, we propose a prediction method of the electric power demand pattern by combining SMO classification techniques and a dimension reduction conceptualized subspace clustering techniques suitable for high-dimensional data cluster analysis. In terms of electricity demand pattern prediction, hourly electricity load patterns and the demographic and geographic characteristics can be analyzed by integrating the wireless load monitoring data as well as sub-regional unit of census information. There are composed of a total of 18 characteristics clusters in the prediction result for the sub-regional demand pattern by using census information and power load of Seoul metropolitan area. The power demand pattern prediction accuracy was approximately 85%.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.3
• /
• pp.813-829
• /
• 2022

In multi-view subspace clustering, how to integrate the complementary information between perspectives to construct a unified representation is a critical problem. In the existing works, the unified representation is usually constructed in the original data space. However, when the data representation in each view is very diverse, the unified representation derived directly in the original data domain may lead to a huge information loss. To address this issue, different to the existing works, inspired by the latest revelation that the data across all perspectives have a very similar or close spectral block structure, we try to construct the unified representation in the spectral embedding domain. In this way, the complementary information across all perspectives can be fused into a unified representation with little information loss, since the spectral block structure from all views shares high consistency. In addition, to capture the global structure of data on each view with high accuracy and robustness both, we propose a novel low-rank approximation via the tight lower bound on the rank function. Finally, experimental results prove that, the proposed method has the effectiveness and robustness at the same time, compared with the state-of-art approaches.