• Journal of KIISE
• /
• v.43 no.8
• /
• pp.910-918
• /
• 2016

A tensor is a multi-dimensional array that represents many data such as (user, user, time) in the social network system. A tensor generator is an important tool for multi-dimensional data mining research with various applications including simulation, multi-dimensional data modeling/understanding, and sampling/extrapolation. However, existing tensor generators cannot generate sparse tensors like real-world tensors that obey power law. In addition, they have limitations such as tensor sizes that can be processed and additional time required to upload generated tensor to distributed systems for further analysis. In this study, we propose TeT, a distributed tera-scale tensor generator to solve these problems. TeT generates sparse random tensor as well as sparse R-MAT and Kronecker tensor without any limitation on tensor sizes. In addition, a TeT-generated tensor is immediately ready for further tensor analysis on the same distributed system. The careful design of TeT facilitates nearly linear scalability on the number of machines.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.1
• /
• pp.13-20
• /
• 2001

Adaptive digital filters with long impulse response such as acoustic echo canceller and active noise controller suffer from slow convergence and computational burden. Subband techniques and multirate signal processing have been recently developed to improve the problem of computational complexity and slow convergence in conventional adaptive filter. Any FIR transfer function can be realized as a serial connection of interpolators followed by subfilters with a sparse impulse response. In this case, each interpolator which is related to the column vector of Hadamard matrix has band-pass magnitude response characteristics shifted uniformly. Subband technique using Hadamard transform and decimation of subband signal to reduce sampling rate are adapted to system modeling and acoustic noise cancellation In this paper, delayless subband structure with nonuniform bandwidth has been proposed to improve the performance of the convergence speed without aliasing due to decimation, where input signal is split into subband one using tree-structured filter bank, and the subband signal is decimated by a decimator to reduce the sampling rate in each channel, then subfilter with sparse impulse response is transformed to full band adaptive filter coefficient using Hadamard transform. It is shown by computer simulations that the proposed method can be adapted to general adaptive filtering.

• Journal of the Korean Society for Industrial and Applied Mathematics
• /
• v.5 no.2
• /
• pp.63-69
• /
• 2001

In this study, we shall be concerned with computing in parallel a few of the smallest eigenvalues and their corresponding eigenvectors of the eigenvalue problem, $Ax={\lambda}Bx$, where A is symmetric, and B is symmetric positive definite. Both A and B are large and sparse. Recently iterative algorithms based on the optimization of the Rayleigh quotient have been developed, and CG scheme for the optimization of the Rayleigh quotient has been proven a very attractive and promising technique for large sparse eigenproblems for small extreme eigenvalues. As in the case of a system of linear equations, successful application of the CG scheme to eigenproblems depends also upon the preconditioning techniques. A proper choice of the preconditioner significantly improves the convergence of the CG scheme. The idea underlying the present work is a parallel computation of the Multi-Color Block SSOR preconditioning for the CG optimization of the Rayleigh quotient together with deflation techniques. Multi-Coloring is a simple technique to obatin the parallelism of order n, where n is the dimension of the matrix. Block SSOR is a symmetric preconditioner which is expected to minimize the interprocessor communication due to the blocking. We implemented the results on the CRAY-T3E with 128 nodes. The MPI(Message Passing Interface) library was adopted for the interprocessor communications. The test problems were drawn from the discretizations of partial differential equations by finite difference methods.

• Proceedings of the Korean Information Science Society Conference
• /
• 2006.10b
• /
• pp.324-329
• /
• 2006

Non-negative matrix factorization(NMF) 기법은 음이 아닌 값으로 구성된 데이터를 두 종류의 양의 행렬의 곱의 형식으로 분할하는 데이터 분석기법으로서, 텍스트마이닝, 바이오인포매틱스, 멀티미디어 데이터 분석 등에 활용되었다. 본 연구에서는 기본 NMF 기법에 기반하여 텍스트 문서로부터 토픽을 추출하고 동시에 이를 가시적으로 도시하기 위한 Topographic NMF (TNMF) 기법을 제안한다. TNMF에 의한 토픽 가시화는 데이터를 전체적인 관점에서 보다 직관적으로 파악하는데 도움이 될 수 있다. TNMF는 생성모델 관점에서 볼 때, 2개의 은닉층을 갖는 계층적 모델로 표현할 수 있으며, 상위 은닉층에서 하위 은닉층으로의 연결은 토픽공간상에서 토픽간의 전이확률 또는 이웃함수를 정의한다. TNMF에서의 학습은 전이확률값의 연속적 스케줄링 과정 속에서 반복적 파리미터 갱신 과정을 통해 학습이 이루어지는데, 파라미터 갱신은 기본 NMF 기반 학습 과정으로부터 유사한 형태로 유도될 수 있음을 보인다. 추가적으로 Probabilistic LSA에 기초한 토픽 가시화 기법 및 희소(sparse)한 해(解) 도출을 목적으로 한 non-smooth NMF 기법과의 연관성을 분석, 제시한다. NIPS 학회 논문 데이터에 대한 실험을 통해 제안된 방법론이 문서 내에 내재된 토픽들을 효과적으로 가시화 할 수 있음을 제시한다.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.12 no.5
• /
• pp.2082-2102
• /
• 2018

In recent years, social network related applications such as WeChat, Facebook, Twitter and so on, have attracted hundreds of millions of people to share their experience, plan or organize, and attend social events with friends. In these operations, plenty of valuable information is accumulated, which makes an innovative approach to explore users' preference and overcome challenges in traditional recommender systems. Based on the study of the existing social network recommendation methods, we find there is an abundant information that can be incorporated into probability matrix factorization (PMF) model to handle challenges such as data sparsity in many recommender systems. Therefore, the research put forward a unified social network recommendation framework that combine tags, trust between users, ratings with PMF. The uniformed method is based on three existing recommendation models (SoRecUser, SoRecItem and SoRec), and the complexity analysis indicates that our approach has good effectiveness and can be applied to large-scale datasets. Furthermore, experimental results on publicly available Last.fm dataset show that our method outperforms the existing state-of-art social network recommendation approaches, measured by MAE and MRSE in different data sparse conditions.

• Quantitative Bio-Science
• /
• v.37 no.2
• /
• pp.133-141
• /
• 2018

Graphical lasso is one of the most popular methods to estimate a sparse precision matrix, which is an inverse of a covariance matrix. The objective function of graphical lasso imposes an ${\ell}_1$-penalty on the (vectorized) precision matrix, where a tuning parameter controls the strength of the penalization. The selection of the tuning parameter is practically and theoretically important since the performance of the estimation depends on an appropriate choice of tuning parameter. While information criteria (e.g. AIC, BIC, or extended BIC) have been widely used, they require an asymptotically unbiased estimator to select optimal tuning parameter. Thus, the biasedness of the ${\ell}_1$-regularized estimate in the graphical lasso may lead to a suboptimal tuning. In this paper, we propose a two-staged bias-correction procedure for the graphical lasso, where the first stage runs the usual graphical lasso and the second stage reruns the procedure with an additional constraint that zero estimates at the first stage remain zero. Our simulation and real data example show that the proposed bias correction improved on both edge recovery and estimation error compared to the single-staged graphical lasso.

• Journal of Information Processing Systems
• /
• v.17 no.3
• /
• pp.462-472
• /
• 2021

With an increase in the scale of recommender systems, users' rating data tend to be extremely sparse. Some methods have been utilized to alleviate this problem; nevertheless, it has not been satisfactorily solved yet. Therefore, we propose an effective pre-rating method based on users' dichotomous preferences and average ratings fusion. First, based on a user-item ratings matrix, a new user-item preference matrix was constructed to analyze and model user preferences. The items were then divided into two categories based on a parameterized dynamic threshold. The missing ratings for items that the user was not interested in were directly filled with the lowest user rating; otherwise, fusion ratings were utilized to fill the missing ratings. Further, an optimized parameter λ was introduced to adjust their weights. Finally, we verified our method on a standard dataset. The experimental results show that our method can effectively reduce the prediction error and improve the recommendation quality. As for its application, our method is effective, but not complicated.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• v.40 no.9
• /
• pp.597-604
• /
• 2016

A parallel algorithm of bi-conjugate gradient method was developed based on CUDA for parallel computation of the incompressible Navier-Stokes equations. The governing equations were discretized using splitting P2P1 finite element method. Asymmetric stenotic flow problem was solved to validate the proposed algorithm, and then the parallel performance of the GPU was examined by measuring the elapsed times. Further, the GPU performance for sparse matrix-vector multiplication was also investigated with a matrix of fluid-structure interaction problem. A kernel was generated to simultaneously compute the inner product of each row of sparse matrix and a vector. In addition, the kernel was optimized to improve the performance by using both parallel reduction and memory coalescing. In the kernel construction, the effect of warp on the parallel performance of the present CUDA was also examined. The present GPU computation was more than 7 times faster than the single CPU by double precision.

• Journal of the Korean Institute of Telematics and Electronics
• /
• v.12 no.1
• /
• pp.27-33
• /
• 1975

The required minimum size of character diode matrix of Korean letters is estimated from the topological complexity of letter structure. The OR aombination of three letter boards (diode matrice) gives all possible Hangout whole letters in proper traditional combined form with minimum required discernibility. Two forms of first consonants (centre located ones for horizontal vowels and leftward displaced ones for vertical and composed vowels) are switched by only 1 bit of the vowel code. The vowel pattern length is modified by again the last four bits of the code. A new 15bit sized inner code is proposed which permits considerably small sized decoding mechanism.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.37 no.6
• /
• pp.48-59
• /
• 2000

Eigendecomposition based direction-of-arrival(DOA) estimation algorithm such as MUSIC(multiple signal classification) is known to perform well and provide high resolution in white noise environment. However, its performance degrades severely when the noise process is not white. In this paper we consider the DOA estimation problem in a colored noise environment as a problem of extracting periodic signals from noise, and we take the problem to the wavelet domain. Covariance matrix of multiscale components which are obtained by taking wavelet decomposition on the noise has a special structure which can be approximated with a banded sparse matrix. Compared with noise the correlation between multiscale components of narrowband signal decays slowly, hence the covariance matrix does not have a banded structure. Based on this fact we propose a DOA estimation algorithm that transforms the covariance matrix into wavelet domain and removes noise components located in specific bands. Simulations have been carried out to analyze the proposed algorithm in colored noise processes with various correlation properties.