Earticle

추천 시스템의 성능을 향상시키기 위하여는 가능한 정확한 사용자 정보를 활용할 필요가 있다. 사용자가 남긴 제품 혹은 서비스 평점은 편리하다는 장점은 있으나, 사용자의 선호를 정확하게 표현하지 못하는 한계가 존재하기 때문에 평점만을 사용한 전통적인 방법의 추천 시스템은 성능의 한계가 존재한다. 따라서 사용자의 선호를 자세하게 파악하기 위해선 평점보다 더욱 자세한 선호가 담긴 리뷰 텍스트를 활용할 수 있다. 하지만 리뷰 텍스트와 평점을 동시에 활용할 경우, 평점은 긍정적인 평점을 남겼지만 리뷰 텍스트에는 부정적인 내용이 담긴 불일치가 존재하는 경우가 있다. 이러한 불일치는 추천시스템의 성능을 저하시킨다고 알려져 있다. 따라서 본 연구는 사용자 리뷰 중 텍스트 리뷰와 평점 간에 차이가 있는 불일치 리뷰를 식별하고 식별된 불일치 리뷰에 기반한 조정된 평점을 이용하여 추천 시스템의 성능을 향상시키고자 한다. 이를 위해 TripAdvisor의 미국 9개 도시(애너하임, 애틀랜타, 시카고, 라스베이거스, 로스앤젤레스, 뉴욕, 올랜도, 필라델피아, 워싱턴)의 2002년부터 2021년까지 1,987개 호텔에 남긴 총 41,810명 고객의 리뷰 데이터 301,346개를 사용하였다. 평점과 리뷰 텍스트의 불일치 리뷰를 식별하기 위해 5개의 Lexicon기반 감성 모델인 Vader, TextBlob, Afinn, Stanza, Bing을 활용한다. 기존 연구에 따라서 사용자 평점과 5개의 모델 중 불일치가 1개라도 있는 경우를 불일치 리뷰로 분류하였다. 불일치 리뷰는 전체 리뷰의 20.99%로 나타났으며 Doc2vec 모델을 활용하여 가장 유사한 리뷰의 평점을 통해 조정된 평점을 부여하였다. Matrix Factorization을 활용하여 기존 평점을 사용한 Original Group, 조정된 평점을 사용한 Adjusted Group, 불일치 리뷰를 제거한 Consistent Group의 추천 시스템 성능을 비교하였다. 그 결과 Adjusted Group의 RMSE는 0.4326으로 Original Group의 RMSE인 0.5218보다 개선된 것을 확인하였다. 또한 벤치마크 시스템인 Consistent Group의 RMSE인 0.3869와 큰 차이를 보이지 않았다. 이는 기존 추천 시스템보다 불일치 리뷰에 대해 조정된 평점을 사용하였을 경우 추천 시스템의 성능이 향상되었음을 의미한다. 또한 불일치 리뷰를 제거한 벤치마크 시스템과 유사한 성능을 통해 제거 없이 모든 사용자에 대해 우수한 추천 시스템이 구축되었음을 알 수 있다.

Many researchers have been focused on designing beauty product recommendation system for a long time because of increased need of customers for personalized and customized recommendation in beauty product domain. In addition, as the application of the deep neural network technique becomes active recently, various collaborative filtering techniques based on the deep neural network have been introduced. In this context, this study proposes a deep neural network model suitable for beauty product recommendation by applying Neural Collaborative Filtering and Generalized Matrix Factorization (NCF + GMF) to beauty product recommendation. This study also provides an implementation of web API system to commercialize the proposed recommendation model. The overall performance of the NCF + GMF model was the best when the beauty product recommendation problem was defined as the estimation rating score problem and the binary classification problem. The NCF + GMF model showed also high performance in the top N recommendation.

본 연구에서는 초미세먼지(PM2.5)의 배출원 및 기여도를 파악하기 위하여 PMF 수용 모델을 활용하여 분석을 진행하 였다. 측정지점은 주변에 산단지역, 주거지역, 농촌지역이 위치하여 복합적인 영향을 받을 수 있는 대전지역으로 선정하 였다. 겨울(1월)과 여름(5, 6월)기간 동안 측정데이터를 수집하였고 분석하였다. PMF 모델 분석 결과 7가지 Factor 즉, 먼 지, 해염입자, 2차 생성 질산염 및 염산염, 2차 생성 황산염, 산업배출, 석탄연소, 자동차배출로 분류하였다. 계절별 기여 도를 확인한 결과 먼지와 해염입자의 경우 5-6월에 발생한 황사 이벤트로 인해 그 기간 동안 기여도가 높게 나타났다. 2 차 생성물질의 경우 질산염과 염산염은 상온에서 상변화가 이루어지는 특징으로 인해 여름보다 겨울에 더 높게 나타났 으며, 석탄연소의 경우 겨울에 더 높게 나타났다. 이를 통해 PMF 결과가 양호하게 나타난 것으로 판단하였다.

In this study, analysis was conducted using the PMF acceptance model to identify the emission source and contribution of fine particle (PM2.5). The measurement point was selected as the Daejeon area, where industrial complex areas, residential areas, and rural areas are located nearby, and thus can be affected by multiple factors. Measurement data were collected and analyzed during the winter (January) and summer (May, June) periods. As a result of the analysis of the PMF model, it was classified into 7 factors, namely, dust, sea salt particles, secondary nitrate and chloride, secondary sulfate, industrial emissions, coal combustion, and vehicle emissions. As a result of checking the contribution by season, the contribution of dust and sea salt particles was high during that period due to the yellow dust event that occurred in May-June. In the case of secondary products, nitrate and hydrochloride were higher in winter than in summer due to the phase change at room temperature, and in the case of coal combustion, they were higher in winter. Through this, it was judged that the PMF result was good.

The rapid growth of e-commerce platforms demands recommendation systems that balance scalability with representational expressiveness. While Matrix Factorization (MF)-based collaborative filtering offers computational efficiency, it fails to capture high-order contextual relationships among items. Graph Neural Networks (GNNs) address this limitation but incur prohibitive computational costs in large-scale, real-time settings. To resolve this tradeoff, we propose a unified hybrid framework integrating MFbased candidate generation and GNN-based re-ranking within a retrieve–then–rank architecture. The MF stage efficiently generates a candidate set, while the GNN stage refines rankings via a heterogeneous graph incorporating user–item interactions, item attributes, and category structures. We further introduce an availability-aware scoring mechanism that integrates real-time stock information to enhance practical applicability. Extensive experiments on a large-scale e-commerce dataset demonstrate that the proposed framework consistently outperforms baseline methods, effectively mitigates data sparsity, and maintains computational efficiency suitable for real-world deployment.

Purpose: This study introduces a methodological framework for muscle synergy analysis to enhance the understanding of sports rehabilitation researchers regarding its practical applications. This work is intended to provide clinicians with a practical understanding of muscle synergy analysis and promote its effective application in rehabilitation. Methods: Muscle synergy refers to the concept of explaining muscle coordination during movements as a combination of a limited number of modules (synergies), thereby offering a powerful approach for quantifying muscle activation patterns. In this framework, electromyography (EMG) signals are decomposed using a nonnegative matrix factorization algorithm, and the optimal number of synergies is determined based on the “variance accounted for” criterion. Each synergy represents a distinct group of muscles, specifying their activation timing and relative contributions throughout movement. To facilitate comprehension, we present a case study employing simulated two-channel EMG data, assuming a single synergy. Results: The dataset included five different signal patterns reflecting proportional, inverse, and independent relationships, mimicking real-world EMG signals observed in sports and clinical settings. Conclusion: This work is intended to provide clinicians with a practical understanding of muscle synergy analysis and promote its effective application in rehabilitation.

The matrix factorization algorithms such as the matrix factorization technique (MF), singular value decomposition (SVD) and the probability matrix factorization (PMF) and so on, are summarized and compared. Based on the above research work, a kind of improved probability matrix factorization algorithm called MPMF is proposed in this paper. MPMF determines the optimal value of dimension D of both the user feature vector and the item feature vector through experiments. The complexity of the algorithm scales linearly with the number of observations, which can be applied to massive data and has very good scalability. Experimental results show that MPMF can not only achieve higher recommendation accuracy, but also improve the efficiency of the algorithm in sparse and unbalanced data sets compared with other related algorithms.

With the increasing abundance of Web Services across Internet, Quality of Service (QoS)-based service recommendation has become a hot issue. It is necessary to predict the missing values of QoS for service recommendation. Because Web services run on the Internet, their network locations may be anther critical factor for QoS prediction. Although there have existed many works on QoS prediction, few consider the influence of the network locations of users or Web Services. In this paper, we propose a novel collaborative QoS prediction framework with network location-based regularization (NLBR). We first elaborate the popular Matrix Factorization (MF) model for missing values prediction. Then, by taking advantage of the local connectivity between Web services users, we incorporate network location information to identify the neighborhood. We conduct the experiments on a public large-scale real-world QoS dataset, Experiments show that our proposed approaches have the better prediction performance compared with the existed approaches.

Most recommender systems based on collaborative filtering aim to provide recommendations for a user in one domain. But data sparsity is a major problem for collaborative filtering techniques. Recently, many scholars have proposed recommendation models to alleviate the sparsity problem by transferring rating matrix in other domains. But different domains have different rating scales (e.g., rating scale may be 1-5 or 1-10). Simple process for the rating scale does not reflect the real situation. The diversity of rating scales may cause the opposite effect, making the recommendation results more imprecise. In this paper, we propose a transfer model which learning the common feature tags from other domain. This model ignores the difference of rating scales between two domains, and focus on studying the feature tags. Using its own rating values to fill the missing value. We first get the different types of users (items) based on non-negative matrix tri-factorization from auxiliary domain. The process we call the user (item) clustering. Than we can get a BP neural network which can judge the type of user according to user's feature tags by studying the features of different types of users (items). And we classify the user (items) which from target domain by exploiting the trained neural network and the users’ feature tags of target domain. Use the average rating values of the same type of users (items) to fill the missing value of target domain. We perform extensive experiments to show that our proposed model outperforms the state-of-the-art CF methods for the cross-domain recommendation task.

Many methods have been applied in image compression and Non-negative matrix factorization (NMF) is one of some approach which could be applied in image compression. Non-Negative Matrix Factorization (NMF) was a realtively new approach to decompose data into two factors with non-negative entries. This paper shows that the NMF method can be applied in image compression using a model of 2x2 pixels weighted matrix rather than using the model of 200x200 pixels, 10x10 pixels and 4x4 pixels weighted matrix. This paper also shows that the 2x2 Weighted matrix model example of ((65536, 1), (1, 65536)) was the best weighted matrix for NMF image compression. Finally, this research proved that by using the weighted matrix with higher determinant value could gain smaller size compressed image.

Matrix factorization-based approaches have proven to be efficient for recommender systems. However, due to the time complexity in composing recommendations, matrix factorization-based approaches are inefficient in dealing with large scale datasets. In this paper, we present a new similar user index-based matrix factorization approach for large scale recommender systems. Finding similar users is the most time-consuming phase in large scale recommender systems. To reduce time to find the similar users, we propose a similar user index in matrix factorization. This paper describes the index structure and algorithms. Several experiments are performed. The results show that our approach is more efficient in dealing with the large dataset as compared with matrix factorization approach without the similar user index.

As an important tool to help users filter Internet information, recommender system has played a very important role wherever in academia or in industrial area. During the past years, different recommendation approaches based on the social network have been proposed with the rapid development of online social networks. Different from the traditional ones which assume all the users are independent and identically distributed, these approaches follow the intuition that a person’s implicit or explicit social network will affect his behaviors on the Web. In this paper, on the basis of the existing work, we fuse a baseline predictor model with an improved social recommendation model and propose a social recommendation algorithm based on probability matrix factorization. The experimental result shows that our method outperforms the existing approaches in accuracy.

In order to deal with the issue of network attacks and enhance the security of the network environment, intrusion detection is gaining more and more attention all over the world. In this paper, a novel intrusion detection method based on improved triangular matrix factorization is presented. As a type of famous mathematical tool, triangular matrix factorization has a good ability to reduce the large amount of high dimensional data. However, the traditional triangular matrix factorization has its inherent drawbacks such as the difficulty of setting the parameter adaptively, so the model of an improved version of triangular matrix factorization together with its concrete algorithm is proposed in this paper firstly. Then, improved triangular matrix factorization is employed to convert the high dimensional data of the network into low dimensional vectors of several matrices, with which the anomaly detection can be realized. Experimental results indicate that the proposed method is promising, and it does significantly enhance the detection accuracy and computational efficiency compared with other current popular ones.

Disease conditions in the human brain can be detected by using medical image analysis. Content Based Image Retrieval of medical images can be used as an alternative to recognize the medical images of the human brain. In CBIR, feature extraction and recognition methods was an important feature because of medical image was different from the general image. In this study medical images that contain clinical information will be used as feature extraction using a canny edge detection and recognition features using non-negative matrix Factorization (NNMF). The purpose of this paper was to describe the use of canny edge detection and NNMF in CBIR. So CBIR could provide information on brain diseases and abnormalities. The results showed that the detection of disease in the human brain can be done by using both methods with good results if done preprocessing using histogram equalization.

We propose a convolutive non-negative matrix factorization method to improve the intelligibility of speech signal in the context of adverse noise environment. The noise bases are prior learned with Non-negative Matrix Factorization (NMF) algorithm. A modified convolutive NMF with sparse constraint is then derived to extract speech bases from noisy speech. The divergence function is selected as an objective function to get a multiplicative update of speech base and its corresponding weight. The weights of prior learned noise bases are also updated in the update rule. Listening experiments are conducted to assess the intelligibility performance of speech synthesized using the proposed algorithm. Experimental results indicate that the proposed method is very effective to improve the intelligibility of the noisy speech in various noise contexts and it outperforms conventional algorithms.

Transmissibility function (TF) was used in structural state detection which examined the change of the response-only vibration characteristics. The technology of signal correlation was utilized to obtain the TF between two testing points. Multiple TFs under different states were calculated as basic TFs and formed a state matrix. Subsequently, non-negative matrix factorization (NMF) was performed for the state matrix. Then both the basic and testing TFs were projected to the feature subspace derived from NMF so as to obtain the state feature index vectors, respectively. Finally, the Euclidean distance between state feature index vectors was defined as the state indicator. The experimental results indicated that this method can achieve better detection accuracy than that using magnitude indicator. Actually, a result with 100 percent correct detection was achieved when proper rule of dimensionality reduction was selected. This method is essentially a multivariate statistical process monitoring (MSPM) method. It is feasible in vibration-based structural state detection in the situation where the excitation signals are unavailable or inaccessible.

This paper presents a new recognition algorithm for plant pathology images based on the Non-negative Matrix Factorization, the proposed algorithm is combined with optimal wavelet packet basis to recognize patterns and conduct data encoding in the internet of things oriented intelligent agricultural system. The experimental results show that the performance of the proposed recognition algorithm is far better than those of the principal component analysis and linear discriminant analysis, and the recognition rate are improved, on average, about 14.65% and 11.18% higher than the rates of the above algorithms respectively. The presented algorithm is characterized by the fast speed, high calculation accuracy and easy hardware implementation.

After the image decomposition with Non-subsampled Shearlet Transform (NSST), the transform coefficients have a larger redundancy. In order to reduce the redundant information, an image fusion algorithm based on NSST and non-negative matrix factorization (NMF) is introduced. The source images are decomposed with NSST into the low-frequency coefficients and the high-frequency sub-band coefficients. The low-frequency coefficients are fused based on NMF theory. The high-frequency coefficients are fused based on Regional Sum Modified-Laplacian (SML) Maximum. Finally, the inverse NSST is used to reconstruct the final fused image. The proposed algorithm can effectively remove redundant information, extract global features and capture more direction details information of multi-source image. Experiments show that the proposed algorithm has obvious advantages and the fused image quality has been greatly improved. The presented algorithm is superior to other fusion algorithms from the objective parameters.