Earticle

본 연구의 목적은 좀 더 정확한 고객 선호도 예측을 위한 협업 필터링 알고리즘의 예측 성능을 평가하기 위한 것이다. 고객 선호도 예측의 정확도를 비교하기 위하여 이웃 기반의 협업 필터링 알고리즘과 대응평균 알고리즘에 의한 고객 선호도 예측의 MAE를 비교하였다. 예측 알고리즘의 정확성을 분석하기 위하여 MovieLens 1 Million dataset을 이용하여 실험을 하였다. 각 예측 알고리즘에 사용된 유사도 가중치는 일반적으로 이용되는 피어슨 상관계수와 벡터 유사도를 이용하였으며 분석결과 대응평균 알고리즘의 예측 정확도가 이웃 기반의 협업 필터링 알고리즘의 예측 정확도 보다 우수한 것으로 나타났다. 두 알고리즘에 사용된 유사도 가중치인 피어슨 상관계수와 벡터 유사도는 두 고객이 특정 상품에 대하여 공통으로 평가한 선호도 평가치를 이용하여 계산된다. 이때 공통으로 평가한 선호도 평가치의 개수가 적으면 계산된 유사도 가중치가 과대 평가된다. 과대 평가된 유사도 가중치를 보정하여 고객 선호도 예측의 정확도를 높이기 위하여 기존의 연구에서 고려한 공통 평가 영화의 개수 보다 확대된 범위를 적용하였으며 각 예측 방법에 따라 서로 다른 개선 경향을 파악할 수 있었다.

The purpose of this study is to evaluate the performance of collaborative filtering recommender algorithms for better prediction accuracy of the customer's preference. The accuracy of customer's preference prediction is compared through the MAE of neighborhood based collaborative filtering algorithm and correspondence mean algorithm. It is analyzed by using MovieLens 1 Million dataset in order to experiment with the prediction accuracy of the algorithms. For similarity, weight used in both algorithms, commonly, Pearson's correlation coefficient and vector similarity which are used generally were utilized, and as a result of analysis, we show that the accuracy of the customer's preference prediction of correspondence mean algorithm is superior. Pearson's correlation coefficient and vector similarity used in two algorithms are calculated using the preference rating of two customers' co-rated movies, and it shows that similarity weight is overestimated, where the number of co-rated movies is small. Therefore, it is intended to increase the accuracy of customer's preference prediction through expanding the number of the existing co-rated movies.

코로나19로 인하여 배달 빈도수가 높은 힘든 1인가구나 맞벌이가족에게는 식자재의 낭비 및 유통기한 관리의 문제점을 가지고 있다. 이런 문제점을 해결하기 위해 본 논문에서는 사용자가 식재료의 유통기한 을 관리하고 식재료 기반 또는 사용자 행동 양식 기반으로 레시피를 추천 받을 수 있게 해주는 애플리케 이션을 설계한다. 본 애플리케이션은 객체 인식 알고리즘인 YOLO를 이용하여 카메라로 식재료를 인식 하고, 협업 필터링을 이용한 레시피 추천 시스템으로 사용자가 원하는 레시피 정보를 제공한다. 사용자 는 본 애플리케이션을 사용하여 카메라로 식재료를 인식하여 레시피를 검색할 수 있고 사용자의 과거이 력이나 취향에 연관된 레시피를 추천 받을 수 있다.

The accuracy of the traditional online recommendation system, much depends on the collaborative filtering recommendation algorithm, however, recommend system aims to attract the interest of consumers and turn visitors into buyers, rather than accurately predict their score. Online recommendation system is the service version of social filtering process. Most previous studies emphasize the accuracy of the collaborative filtering algorithm. However, the effective recommendation system must be credible. It requires that the system logic be transparency and the system be able to provide consumers a new, inexperienced item. Based on the above, this paper proposes to research the quality evaluation of recommendation system from the angle of user’s experience, adding a freshness parameters of Top-N recommend collaborative filtering similarity calculation method, and comparing with the classical recommended algorithm. The experiment result has a certain degree of accuracy and high diversity, which provides basis for establishing the e-commercial recommendation system.

A new algorithm FTRA has been proposed, which infuses users’ trust network and rating data. The sparse problem of rating data will significantly reduce the accuracy of collaborative filtering recommendation. In addition to the users’ ratings data on the Internet, other data sources which can be used in the process of recommend, and one of the more common is trust network data which describes the mutual relationship between users. To solve this problem, this paper will the data of trust network as an important supplement on the rating data, and bases on graph theory concepts or methods, the similarity method in the paper, and the Katz method which is used to calculate the similarity of link, proposes the FTRA algorithm which organic infuses this two data, and then better to solve the sparse problem of the rating data faced by collaborative filtering. The experimental results on the Epinions dataset show that the FTRA algorithm is superior to or significantly better than the comparison algorithms, which include the algorithms that only based on the rating data or the trust relationship, and the other algorithms infusing the two data sources.

Aiming at the problems that the existing model-based collaborative filtering algorithm has low recommendation accuracy and small recommendation coverage, we propose a collaborative filtering recommendation algorithm based on the trust propagation by introducing the trust information of social network to extend the matrix factorization-based recommendation model. We first design a set of trust propagation rules based on the direct trust relationships of the social network, so as to propagate the trust relationship in the social networks, and get to quantize the new trust relationship. Then we load the quantitative trust relations after the trust propagation as the trust weight into the matrix factorization-based model according to the characteristics that the matrix factorization technique can reduce the dimension of large-scale datasets.

To overcome the problem of too much sparse by scoring matrix, the paper proposes from the part of user interests the collaborative filtering recommendation algorithm based on such interest. In the e-commerce websites, recommendation items have various appearance, function and category attributes. Items which hold similar characteristics generally gain approximate scoring values. The aforesaid method extracts useful information for improvement from user interests. Users’ scorings about different items imply their modes of interest. Interest association exists amongst items that were evaluated by the same user. Since individual interest shifts, the intensity of such correlation will gradually change along with days. By building interest intensity model with time decay, and discovering interest correlation among different items through that model, the proposed algorithm can predict scoring matrix and fill it, which is helpful to alleviate problems with sparse caused by user-item scoring matrix.

Traditional methods UBCF have limitations of poor recommendation quality and problems of data sparsity. To alleviate these problems, a novel collaborative filtering algorithm is designed, which firstly get the users’ ratings and time intervals for each attribute from the users’ ratings for items, then produce two methods to calculate the similarity between users, introduce a weighting parameters to control the weight between the two similarity methods in order to get a fusion similarity between two users. The results show that this method is able to improve the accuracy of predicted values, resulting in improving recommendation quality of the collaborative filtering recommendation algorithm.

The paper Proposed Item parallel collaborative filtering recommendation algorithm (IP-CF). Through designing efficient parallel algorithm, compute-extensive procedures are distributed to different processing nodes in Hadoop platform. Taking advantage of parallel computing, we accelerate the response of recommendation. The experimental results show that our proposed algorithm IP-CF is more efficient and scalable than current parallel algorithms.

An improved CF algorithm based on the classification of items is introduced to overcome the problems caused by the data sparseness and inaccuracy of the user neighbors. The new algorithm first rates the unrated items by applying the item classification, and then calculates the user similarity within classes for nearest-neighbors, after which it could recommend the items based on the final prediction.

Collaborative filtering recommendation technology is by far the most widely used and successful personalized recommendation technology. However, the method currently faced with some problems such as sparse matrix, affecting the accuracy of the predicted results. This paper puts forward a new community detection algorithm based on topological potential theory, and combines it with collaborative filtering recommendation algorithm. The users with similar interests are put into the same community. When searching for the user’s nearest neighbor, it target to the users in a specific community or several communities instead of all users, which narrows the search and improves the prediction accuracy. Experimental results suggest that this approach effectively reduces the impact on the prediction accuracy of the sparse matrix, and significantly improves the prediction ability and recommendation quality.

As an important one of recommendation technologies, collaborative filtering algorithms have many advantages, and have been very successful in both research and practice. However they also remain fundamental challenges, such as data sparsity, cold-start, dynamic changes of users’ preferences and interests, and scalability. For purpose of lessening inaccurate recommendations caused by data sparsity, the pre-filled rating matrix is created by introducing a novel similarity computation method to replace the traditional ones. To address the cold-start issue, we propose a hybrid recommendation method that combines collaborative filtering and content-based filtering exploiting the advantages of both methods. To respond positively to dynamic changes of users’ preferences and interests, the improved algorithm takes time factor into consideration as well. Finally we implement parallel execution of the improved algorithm on Hadoop platform, which addresses serious scalability issue when working on big data. The experimental evaluation of our proposed methods took place and the results showed that the improved algorithm has better recommendation quality and real-time performance.

This paper has improved the accuracy of recommendation by an improved collaborative filtering algorithm based on conventional computing method. By considering the preferences of different user and the similarity of their respective preferences, it proposed a similarity formula based on logarithm. Experimental results show that compared to traditional collaborative filtering algorithms, the new recommendation algorithm dramatically increased the recommendation accuracy, with better personalized recommendation effect.

Aiming at the problem that the traditional Collaborative Filtering algorithm has low recommendation accuracy, in the paper, we propose a collaborative filtering recommendation algorithm based on the global trust degree integrating the direct trust information in the social networks. We first transform the local trust relationships to the global trust relationship by the rules in the trust network, and get the trust rank of all users in the trust networks; Then we use the global trust value to instead of the similarity information value as the weights of a predicted formula in the traditional collaborative recommendation algorithm, and integrate the weights to the matrix factorization-based recommendation model.

Under the conditions of different community formation, this paper proposed two different models of formation communities. Firstly, we put forward two kinds of similarity calculation models, and compare them with the traditional similarity model, Secondly, several similarity models are tested under different conditions of community formation. Finally it compares tow models of forming communities and finds that for non-strict division of community model has a higher accuracy and diversity of recommendation, compared with the strict division of community model. Thus, the experiments show that the non-strictly divided communities’ model is more suitable for recommendation system, especially for the personalized recommendation.

본 논문에서는 협업 필터링 알고리즘을 클러스터링 기반으로 분산 환경에서 구현하여, 추천을 위한 수행 시 간을 최적화 하는 방법에 대한 제안을 한다. 하둡 기반으로 시스템을 구성하였고, 분산 Min-hash 클러스터링 기반의 협업 필터링 방법을 제안하고, 이를 기반으로 분산 추천 시스템을 구성하였다. 분산 사용자 기반 협업 필터링 기법을 사용하여 무비렌즈 (Movie Lens)의 영화 평점 데이터를 기반으로 각각의 사용자에게 알맞은 영화를 추천해주는 분산 추천 시스템을 구현하고 실험을 통하여 성능의 우수성을 검증하였다.

This paper presents an efficient distributed recommendation system using clustering collaborative filtering algorithm in distributed computing environments. The system was built based on Hadoop distributed computing platform, where distributed Min-hash clustering algorithm is combined with user based collaborative filtering algorithm to optimize recommendation performance. Experiments using Movie Lens benchmark data show that the proposed system can reduce the execution time for recommendation compare to sequential system.

우리는 개인 전자상거래 포탈에서 개인화를 위한 시맨틱 추천 방법을 제안한다. 시맨틱 추천은 제품의 특성(속성)을 이용하여 의미적 유사성 평가를 통해 이루어진다. 정확한 추천을 제공하기 위하여 제품의 시맨틱 유사성은 제품의 평점정보를 포함한다. 또한, 추천기술은 제품의 평점을 평가하여 고객의 다양한 내포된 의향을 분석한다. 고객의 의향은 "구입한 제품", "쇼핑카트에 추가한 제품", "정보를 본 제품"과 같이 세 가지 유형으로 분류 하고 있다. 우리는 제품의 추천을 위한 제품의 평점을 추정하기 위하여 고객의 내재적 의향을 추적할 수 있다. 또한 우리는 정확한 추천을 제공하기 위해 매우 중요한 유효한 세션을 식별하는 유효성 검사 프로세스 세션을 구현하였다. 우리의 추천 기술은 유사한 환경의 고객의 연령별 그룹에서 높은 수준을 정확도를 보여 준다. 본 논문의 실험섹션에서 우리의 제안 추천방식은 기존 고객뿐만 아니라 이전의 구매기록이 없는 새로운 사용자에게도 기존에 잘 알려진 협업 필터링 방법보다 좋은 성능을 보여 주었다.

This paper proposes a semantic recommendation technique for a personalized e-commerce portal. Semantic recommendation is achieved by utilizing the attributes of products. The semantic similarity of the products is merged with the rating information of the products to provide an accurate recommendation. The recommendation technique also analyzes various attitudes of the customer to evaluate the implicit rating of products. Attitudes are classifies into three types such as "purchasing product", "adding product to shopping cart", and "viewing the product information." We implicitly track customer attitude to estimate the rating of products for recommending products. Also we implement a session validation process to identify the valid sessions that are highly important for giving an accurate recommendation. Our recommendation technique shows a high degree of accuracy as we use age groupings of customers with similar preferences. The experimental section shows that our proposed recommendation method outperforms well known collaborative filtering methods not only for the existing customer, but also for the new user with no previous purchase record.

최근 사용자들의 추천 서비스를 위해 다른 사용자들의 평가값을 이용하여 특정 사용자에게 서비스를 추천해주는 추천 시스템은 협력 필터링 방법을 널리 사용되고 있다. 하지만 이러한 추천 시스템은 클러스터링 과정에서 이미 분류된 그룹에 특정 사용자가 분류되어 정확히 분류되지 못하고, 사용자들의 평가값 오차가 클 경우 정확하지 못한 결과를 추천하는 문제점이 있다. 본 논문에서는 이러한 문제점을 해결하기 위하여 협력 필터링 알고리즘을 클러스터링 기반으로 분산 환경에서 구현하여, 추천의 효과를 최적화 하는 기법을 제안하며 하둡 및 스파크 기반으로 시스템을 구성하여 협력 필터링 추천 알고리즘을 비교 하였다.

추천 시스템을 위한 분석방법들 가운데 협업 필터링은 데이터 분석에 기반한 추천 시스템에서 주요 대표적 방법이다. 일반적 사용 방법은 다양한 아이템에 대해서 사용자들의 평가 데이터를 활용하여 공통적인 패턴을 찾으며, 특정 사용자에 대한 선호 아이템을 추천하는 기법이다. 이에 본 논문에서는 여러가지 알고리즘을 사용하여 지표 측정에 활용하였으며, 사용자 선호에 대한 예측에 적합한 알고리즘을 찾아서 제시하였다.

Among the analysis methods for a recommendation system, collaborative filtering is a major representative method in a recommendation system based on data analysis. A general usage method is a technique of finding a common pattern by using evaluation data of users for various items, and recommending a preferred item for a specific user. Therefore, in this paper, various algorithms were used to measure the index, and an algorithm suitable for prediction of user preference was found and presented.

본 논문에서는 고객의 특성을 고려한 최적의 추천시스템을 개발하기 위하여 기존의 인구통계학적 특성에 따른 협동적필터링 기법의 추천 효율을 비교 분석하였다. 비디오에 대한 사용자 평가 값과 예측 값간의 추천 효율에 대한 비교실험을 통하여 상품에 대한 단순한 선호도만을 고려한 기존의 협동적필터링 방법에 의한 추천시스템의 문제점을 개선하여 추천된 상품이나 콘텐츠에 대한 개인별 추천 효율을 향상시키기 위한 모델을 제시하였다. 본 연구 결과를 이용하여 인터넷 비즈니스 분야에서 활발하게 도입되고 있는 eCRM 시스템에서 가장 중요한 요소인 고객들의 인구통계학적인 다양한 특성을 고려한 협동적필터링 기반의 추천시스템을 개발할 수 있으리라 기대한다.