• Journal of Information Technology Applications and Management
• /
• v.16 no.1
• /
• pp.97-113
• /
• 2009

Collaborative filtering(CF) has been a very successful approach for building recommender system, but its widespread use has exposed to some well-known problems including sparsity and scalability problems. In order to mitigate these problems, we propose two novel models for improving the typical CF algorithm, whose names are ISCF(Item-Selected CF) and USCF(User-Selected CF). The modified models of the conventional CF method that condense the original dataset by reducing a dimension of items or users in the user-item matrix may improve the prediction accuracy as well as the efficiency of the conventional CF algorithm. As a tool to optimize the reduction of a user-item matrix, our study proposes genetic algorithms. We believe that our approach may relieve the sparsity and scalability problems. To validate the applicability of ISCF and USCF, we applied them to the MovieLens dataset. Experimental results showed that both the efficiency and the accuracy were enhanced in our proposed models.

• Journal of Information Technology Applications and Management
• /
• v.12 no.3
• /
• pp.41-56
• /
• 2005

Product recommender system is one of the most popular techniques for customer relationship management. In addition, collaborative filtering (CF) has been known to be one of the most successful recommendation techniques in product recommender systems. However, CF has some limitations such as sparsity and scalability problems. This study proposes hybrid cluster analysis and case-based reasoning (CBR) to address these problems. CBR may relieve the sparsity problem because it recommends products using customer profile and transaction data, but it may still give rise to scalability problem. Thus, this study uses cluster analysis to reduce search space prior to CBR for scalability Problem. For cluster analysis, this study employs hybrid genetic and K-Means algorithms to avoid possibility of convergence in local minima of typical cluster analyses. This study also develops a Web-based prototype system to test the superiority of the proposed model.

• Journal of Broadcast Engineering
• /
• v.26 no.7
• /
• pp.862-867
• /
• 2021

Although Deep Neural Networks (DNNs) have shown remarkable performance in various artificial intelligence fields, it is well known that DNNs are vulnerable to adversarial attacks. Since adversarial attacks are implemented by adding perturbations onto benign examples, increasing the sparsity of DNNs minimizes the propagation of errors to high-level layers. In this paper, unlike the traditional pruning scheme removing low magnitude weights, we eliminate high magnitude weights that are usually considered high absolute values, named 'reverse pruning' to ensure robustness. By conducting both theoretical and experimental analyses, we observe that reverse pruning ensures the robustness of DNNs. Experimental results show that our reverse pruning outperforms previous work with 29.01% in Top-1 accuracy on perturbed CIFAR-10. However, reverse pruning does not guarantee benign samples. To relax this problem, we further conducted experiments by adding a regularization term for the high magnitude weights. With adding the regularization term, we also applied conventional pruning to ensure the robustness of DNNs.

• Journal of information and communication convergence engineering
• /
• v.17 no.2
• /
• pp.135-141
• /
• 2019

Collaborative filtering algorithms often encounter data sparsity issues. To overcome this issue, auxiliary information of relevant items is analyzed and an item attribute matrix is derived. In this study, we combine the user-item attribute preference with the traditional similarity calculation method to develop an improved similarity calculation approach and use weights to control the importance of these two elements. A collaborative filtering algorithm based on user-item attribute preference is proposed. The experimental results show that the performance of the recommender system is the most optimal when the weight of traditional similarity is equal to that of user-item attribute preference similarity. Although the rating-matrix is sparse, better recommendation results can be obtained by adding a suitable proportion of user-item attribute preference similarity. Moreover, the mean absolute error of the proposed approach is less than that of two traditional collaborative filtering algorithms.

• Journal of the Korea Society of Computer and Information
• /
• v.23 no.12
• /
• pp.219-226
• /
• 2018

Collaborative filtering has been most popular approach to recommend items in online recommender systems. However, collaborative filtering is known to suffer from data sparsity problem. As a simple way to overcome this problem in literature, Jaccard index has been adopted to combine with the existing similarity measures. We analyze performance of such combination in various data environments. We also find optimal weights of factors in the combination using a genetic algorithm to formulate a similarity measure. Furthermore, optimal weights are searched for each user independently, in order to reflect each user's different rating behavior. Performance of the resulting personalized similarity measure is examined using two datasets with different data characteristics. It presents overall superiority to previous measures in terms of recommendation and prediction qualities regardless of the characteristics of the data environment.

• Journal of Information Processing Systems
• /
• v.15 no.2
• /
• pp.410-421
• /
• 2019

Distributed compressed sensing (DCS) states that we can recover the sparse signals from very few linear measurements. Various studies about DCS have been carried out recently. In many practical applications, there is no prior information except for standard sparsity on signals. The typical example is the sparse signals have block-sparse structures whose non-zero coefficients occurring in clusters, while the cluster pattern is usually unavailable as the prior information. To discuss this issue, a new algorithm, called backtracking-based adaptive orthogonal matching pursuit for block distributed compressed sensing (DCSBBAOMP), is proposed. In contrast to existing block methods which consider the single-channel signal reconstruction, the DCSBBAOMP resorts to the multi-channel signals reconstruction. Moreover, this algorithm is an iterative approach, which consists of forward selection and backward removal stages in each iteration. An advantage of this method is that perfect reconstruction performance can be achieved without prior information on the block-sparsity structure. Numerical experiments are provided to illustrate the desirable performance of the proposed method.

• ETRI Journal
• /
• v.41 no.5
• /
• pp.658-669
• /
• 2019

For dual-channel time-frequency (TF) overlapped signals with low sparsity in underdetermined blind source separation (UBSS), this paper proposes an effective method based on interval probability to estimate and expand the types of mixing matrices. First, the detection of TF single-source points (TF-SSP) is used to improve the TF sparsity of each source. For more distinguishability, as the ratios of the coefficients from different columns of the mixing matrix are close, a local peak-detection mechanism based on interval probability (LPIP) is proposed. LPIP utilizes uniform subintervals to optimize and classify the TF coefficient ratios of the detected TF-SSP effectively in the case of a high level of TF overlap among sources and reduces the TF interference points and redundant signal features greatly to enhance the estimation accuracy. The simulation results show that under both noiseless and noisy cases, the proposed method performs better than the selected mainstream traditional methods, has good robustness, and has low algorithm complexity.

• ETRI Journal
• /
• v.45 no.3
• /
• pp.492-504
• /
• 2023

Researchers have recently shown an increased interest in estimating the direction-of-arrival (DOA) of wideband noncircular sources, but existing studies have been restricted to subspace-based methods. An off-grid sparse recovery-based algorithm is proposed in this paper to improve the accuracy of existing algorithms in low signal-to-noise ratio situations. The covariance and pseudo covariance matrices can be jointly represented subject to block sparsity constraints by taking advantage of the joint sparsity between signal components and bias. Furthermore, the estimation problem is transformed into a single measurement vector problem utilizing the focused operation, resulting in a significant reduction in computational complexity. The proposed algorithm's error threshold and the Cramer-Rao bound for wideband noncircular DOA estimation are deduced in detail. The proposed algorithm's effectiveness and feasibility are demonstrated by simulation results.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.24-26
• /
• 2022

시스톨릭 어레이는 DNN training 등 인공지능 연산의 대부분을 차지하는 행렬 곱셈을 수행하기 위한 하드웨어 구조로 많이 사용되지만, sparsity 가 높은 행렬을 연산할 때 불필요한 동작으로 인해 효율성이 크게 떨어진다. 본 논문에서 제안된 유동적 시스톨릭 어레이는 matrix condensing, weight switching, 그리고 direct output path 의 방법과 구조를 통해 sparsity 가 높은 행렬 곱셈의 수행 사이클을 줄일 수 있다. 시뮬레이션을 통해 기존 시스톨릭 어레이와 유동적 시스톨릭 어레이의 성능을 비교하였으며 8×8, 16×16, 32×32 의 크기를 가진 행렬을 동일 크기의 시스톨릭 어레이로 연산하였을 때 필요 사이클 수를 최대 12 사이클 절감할 수 있는 것을 확인하였다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.437-442
• /
• 2023

최근에 등장한 대규모 언어 모델은 다양한 언어 처리 작업에서 놀라운 성능을 발휘하고 있다. 그러나 이러한 모델의 크기와 복잡성 때문에 모델 경량화의 필요성이 대두되고 있다. Pruning은 이러한 경량화 전략 중 하나로, 모델의 가중치나 연결의 일부를 제거하여 크기를 줄이면서도 동시에 성능을 최적화하는 방법을 제시한다. 본 논문에서는 한국어 언어 모델인 Polyglot-Ko에 Wanda[1] 기법을 적용하여 Pruning 작업을 수행하였다. 그리고 이를 통해 가중치가 제거된 모델의 Perplexity, Zero-shot 성능, 그리고 Fine-tuning 후의 성능을 분석하였다. 실험 결과, Wanda-50%, 4:8 Sparsity 패턴, 2:4 Sparsity 패턴의 순서로 높은 성능을 나타냈으며, 특히 일부 조건에서는 기존의 Dense 모델보다 더 뛰어난 성능을 보였다. 이러한 결과는 오늘날 대규모 언어 모델 중심의 연구에서 Pruning 기법의 효과와 그 중요성을 재확인하는 계기가 되었다.