Elasticsearch 기반의 동적 추천 아키텍쳐 및 절차

정지현,김철진 한국지식정보기술학회 2020 한국지식정보기술학회 논문지 Vol.15 No.4

E-commerce product recommendation uses various recommendation techniques. The recommendation techniques that are generally applied include the collaborative filtering technique, the user based collaborative filtering, and the item based collaborative filtering technique, and utilizes recommended techniques based on artificial intelligence technologies such as Recurrent Netural Network (RNN) and Long Short Term Memory (LSTM). The recommendation data generated through these recommendation techniques are stored in a database and are recommended in the form of static data by user actions (search, order, etc.). However, the recommendation through the static recommendation product list is limited in improving the purchasing power of the user. Providing a variety of recommended products and real-time properties, it is possible to further improve purchasing power if it is possible to dynamically construct a recommended product list. Therefore, this paper proposes an architecture and procedure for providing dynamic recommendations. The proposed dynamic recommendation architecture constructs a dynamic recommendation product by applying a static recommendation product based on Elasticsearch. In the experiment, accuracy was compared and analyzed by applying the same transaction data to existing recommendation techniques, and it was confirmed that the accuracy of the dynamic recommendation technique proposed in this paper is improved compared to the existing recommendation techniques.

데이터 분석 기반 추천학습모델 생성도구 설계 및 구현

정지현,조천우,변동훈,김철진 한국지식정보기술학회 2021 한국지식정보기술학회 논문지 Vol.16 No.1

Current sales of offline products are conducted on a face-to-face recommendation based on customer’s preferences and fields of interest. However, as the online shopping malls are actively spreading to the public, recommendation researches are ongoing to make more accurate recommendation using customers’ online activity data besides a face-to-face recommendation. But compared to recommendation systems of large-scale shopping mall where operates big-scale database and professional human resources, small-scale shopping malls use simple recommenders relatively, which can be interpreted as a judgement that those recommenders developed by large-scale companies are unfit to apply onto their system. In this paper, we propose a platform that can develop a generalized recommendation tool by utilizing Open API. Recommender-development tool of this study refines datasets to be used for developing a recommendation system, analyzes the datasets to derive proper algorithms. After doing those progresses, it proceeds model learning progress based on the derived algorithm and provides Open API to a user so that the user can apply the recommendation system to own shopping mall system. We prove proposed tool by developing recommendation system using real transaction datasets with this tool, lastly mounting the Open API onto the test shopping mall site. While users using existing tools must be directly specified the schema of datasets to make a recommendation system, this study has a distinction in that it allows users to develop the recommendation model by automatically proceeding with this process.

THE ASYMMETRIC EFFECTS OF ATTITUDE TOWARD THE BRAND (SYMBOLIC vs. FUNCTIONAL) UPON RECOMMENDATION SYSTEM (ARTIFICIAL INTELLIGENCE vs. HUMAN)

Kiwan Park,Yaeri Kim,Seojin Stacey Lee 글로벌지식마케팅경영학회 2018 Global Marketing Conference Vol.2018 No.07

New product entails risk, causing resistance to adoption. The recommendation system may decrease the psychological risk by guiding decision making process to be more efficient (H?ubl and Trifts, 2000). AI (Artificial Intelligence) has been getting smarter and smarter and widely applied to the recommendation system. Even while you are browsing on your Facebook, AI recommends you the products that you may like based on the customized analysis of your interest. However, do people always love to adopt the smart recommends from AI? Definitely no! Then when and why people reluctantly accept AI recommendation? We assume that the product or service where the sense and feeling is important, people might be reluctant to accept the recommendation from artificial intelligence. This is because people might feel threatened when the AI challenges against human uniqueness (Gray and Wegner, 2012). Thus, in this study we investigated how the recommendation system types (AI vs. Human) affect brand attitude depending on the brand image (Symbolic vs. Functional). We found consumers are reluctant to accept a recommendation from AI in symbolic brand where human sense and feel are considered to be critical factors (Study1). This effect was further explained by uncanny-feeling toward the AI recommendation system (Study2). This research is meaningful in that it is the first attempt to apply the artificial intelligence recommendation concept to the marketing strategy by incorporating the concept of brand image. We predicted and found AI based recommendation system is reluctantly accepted for symbolic brand. Furthermore, we discovered the underlying process for this phenomenon as uncanny feeling. People seemed to have uncomfortable feelings against AI recommendation when the brand image is related to sense and feel considered as nature of human uniqueness. Thus, marketers should be very cautious when utilizing the AI recommendation system not to threaten human uniqueness area.

설명기능을 추가한 협업필터링 기반 개인별 상품추천시스템

조윤호(Yoon Ho Cho),김재경(Jae Kyeong Kim),안도현(Do Hyun Ahn),이희애(Hee Ae Lee) 한국경영학회 2006 經營學硏究 Vol.35 No.2

The continuous growth of the Internet and e-commerce has allowed companies to provide customers with more choices on products. Increasing choice has also caused product overload where the customer is no longer able to effectively choose the products he/she is exposed to. A promising technology to overcome the product overload problem is recommender systems that help customers find the products they would like to purchase. To date, a variety of recommendation techniques have been developed. Collaborative Filtering (CF) is the most successful recommendation technique, which has been used in a number of different applications such as recommending movies, articles, books, Web pages, etc. However, its widespread use has exposed some limitations, such as sparsity, scalability, and black box. Many researchers have focused on sparsity and scalability problems but a little has tried to solve the black box problem. Most CF recommender systems are black boxes, providing no transparency into the working of the recommendation. To overcome the black box problem, it is developed a recommender system named WebCF-Exp (Web usage mining driven Collaborative Filtering with Explanation facilities). Explanation facilities make it possible to expose the reasoning and data behind a recommendation. Therefore, they provide us with a mechanism for handling errors that come with a recommendation. Furthermore, it is proposed to use web usage mining and product taxonomy to enhance the recommendation quality for e-commerce environment. Web usage mining populates the rating database by tracking customers’shopping behaviors in the Web, thereby leading to better recommendations. The product taxonomy is used to improve the performance through dimensionality reduction of the rating database.The overall procedure of WebCF-Exp consists of two phases: recommendation phase and explanation phase. The recommendation phase is divided into four sub-phases: grain specification, customer profile creation, neighborhood formation, and recommendation generation. The explanation phase consists of white box model and black box model. A white box model is one way to build explanation interfaces using detailed process or data such as neighbor ratings, the ratio of click, and the ratio of purchase. Black box model is the other way of which there is no detailed process or data available. In black box model, we focus on ways to justify recommendation that are independent of the process such as marketing or event information.WebCF-Exp recommender system is operated by four agents: Web log analysis agent, Data transformation agent, Recommendation agent, and Explanation agent. Web log analysis agent manages Web log database through periodic collecting, parsing and analyzing Web server log files such as access logs, referrer logs, agent logs and cookie files. Thus, the users can easily access and analyze them like other operation databases. Data transformation agent creates and manages the data mart that provides data indispensable to accomplish recommendation tasks. Recommendation agent makes a personalized recommendation list for each target customer. Explanation agent provides interfaces which expose the reasoning and data behind a recommendation. Twenty different explanation interfaces are developed as white box model and black box model. To test the performance of WebCF-Exp, it is developed a prototype internet shopping mall named EBIB (e-Business & Intelligent Business) and interactive interfaces. Experiments are conducted with the data provided by EBIB Research Internet shopping mall. Our experimental result shows that WebCF-Exp recommendation system shows better performance than the CF recommendation system without explanation facilities. And explanation types of five stars, simple graph, and showing the evaluation results of similar customers, show better performance than other types. Furthermore, as customers understand explanation interfaces better, it results that customers

다수의 대중추천인가? 소수의 지인추천인가?

심선영(Seonyoung Shim) 한국전자거래학회 2012 한국전자거래학회지 Vol.17 No.3

최근 SNS상에서 소비자들이 지인의 추천을 통해 구매를 하는 현상이 많이 벌어지고 있다. 본 연구에서는 이러한 지인추천이 새로운 구매 휴리스틱 유발 요소로서 영향력을 갖는지 살펴본다. 이를 위해 구매 휴리스틱에 대한 지인추천의 영향력을 대표적 휴리스틱 유발 요소인 대중추천의 영향력과 신뢰성, 전문성, 적합성 측면에서 비교해 본다. 나아가 정보원천의 영향력뿐만 아니라 정보 빈도의 효과도 살펴본다. 즉, 지인의 추천이나 대중의 추천이 구매의사결정에 영향을 미침에 있어 정보원천으로서의 우위가 있다면, 그 효과가 상대적 빈도에 의해서는 어떻게 달라지는지도 살펴보는 것이다. 이는, 다수의 대중추천보다는 한정된 지인추천이 양적인 열세를 가질 수 있다는 현실에 착안한 것이다. 따라서 지인추천이라는 새로운 정보원천이 가지는 구매 휴리스틱 영향력에 있어 빈도의 제한성에서 오는 현실적 효과를 살피고자 한다. 연구 결과, 동일한 빈도에서는, 지인추천이 대중추천보다 구매 휴리스틱 유발에 있어 월등한 효과를 가지고 있지 않은 것으로 나타났다. 하지만, 지인추천 또한 대중추천처럼, 아무런 추천이 없는 경우에 비해서는 구매 설득력이 우월한 것이 확인되었으며, 신뢰성 면에서는 대중추천보다 높이 평가되었다. 또한, 지인추천이 대중추천보다 강력한 구매 휴리스틱 요소가 되기 위해서는 절대적 빈도 우위가 필요함도 밝혀졌다. 본 연구는 소비자의 구매 휴리스틱에 대한 이해를 넓힘으로써, 구매에 보다 적절한 정보를 제공하고, 효율적인 구매를 지원할 수 있도록 기업관점의 함의를 제공해 줄 것이다. Recently, there happens many purchasing cases encouraged by friends recommendation in SNS (Social Network Service). This study investigates the effect of friend recommendation on consumers purchasing heuristic. For this purpose, we compare the effect of friend recommendation with consumer recommendation in terms of trustworthy, specialty, relevancy. Usually, the frequency of friend recommendation is far lower than that of consumer recommendation. Hence, we examine how the effect of information source (friend recommendation or consumer recommendation) is moderated by the frequency of recommendation, as well. As results, this study finds out that, under the same frequency, friend recommendation does not have significantly stronger effect on the purchasing heuristic, although friend recommendation is evidenced as one of significant heuristic inducers. However, in terms of trustworthy, friend recommendation is significantly superior to the consumer recommendation. Moreover, under sufficiently higher frequency, friend recommendation works as better heuristic factor than consumer recommendation. The results deliver managerial implications in the perspective of understanding consumers’ purchasing decisions and responding strategies of firms.

추천 알고리즘의 이유 제공 방식에 따른 사용자 경험 연구 -동영상 추천 플랫폼을 중심으로-

조윤지,임영송,송준근,정아윤,김기현,윤 재 영 (사)한국커뮤니케이션디자인협회 커뮤니케이션디자인학회 2023 커뮤니케이션 디자인학연구 Vol.83 No.-

Living in an information society, people tap into state-of-the-art technology, specialized in various fields. For example, the AI algorithms can be used to recommend a variety of contents in some platform services but the way of recommending the unilateral contents is controversial. Thus, this study aims to examine how the user experience varies in accordance with the way of providing a reason for recommendation. Through the case analysis, the way of providing a reason for recommendation was divided into 'no information-type', 'message-type', 'tooltip-type', and 'report-type'. Also, the surveys and interviews were conducted on usefulness, reliability, satisfaction, and intention of continuous use the recommended algorithms. According to the way of providing a reason for recommendation, user experience was different, especially, 'report-type' showed significant differences in all user experience factors compared to the 'no information' type. This study confirmed that it is necessary to explain the outcome of the recommendation algorithm by experimentally verifying the user experience from the way of providing a reason for recommendation. It is essential that the operating function of the recommendation algorithms come together with user's right to know, which is contributing to practical usefulness and enjoyment, suggesting the direction of use based on reliable value. 정보 사회를 살아가는 현대인들은 고도화된 전문 기술을 다양하게 활용하고 있다. 그중 알고리즘은 뉴스, 음악, 영상 스트리밍 등 추천 서비스에 활용되어 사용자에게 선택의 편리함을 제공한다. 그러나 사용자 데이터를 기반으로 콘텐츠가 일방적으로 추천되는 방식에 대해 논란이 되고 있다. 따라서 본 연구에서는 추천 알고리즘의 추천 이유 제공 방식에 따라 변하는 사용자 경험을 살펴보았다. 사례 분석을 통해 추천 이유 제공 방식을 '정보없음형', '메시지형', '툴팁형', '리포트형'으로 구분한 후 유형별 실험물을 제작하여 추천 알고리즘의 이유 제공 방식에 대한 유용성, 신뢰도, 만족도, 지속사용의도에 대한 설문 조사와 인터뷰를 진행하였다. 연구 결과 추천 알고리즘의 이유 제공 방식에 따라 사용자 경험은 차이를 보였으며, 특히 '리포트형'은 현재 추천 서비스와 유사한 '정보없음형'과 비교해 모든 사용자 경험 요인에서 유의미한 차이를 보였다. 본 연구는 추천 알고리즘의 이유 제공 방식에 따른 사용자 경험을 실증적으로 검증하여 추천 알고리즘의 결과를 설명하는 것이 필요함을 확인하였다. 이는 추천 알고리즘의 사용 목적이 기능적 유용성과 즐거움 제공에서 사용자들의 알권리를 충족시켜 신뢰 가치를 바탕으로 한 활용 방향성을 제시했다는 것에 의의가 있다.

OTT 서비스에서 사용자 콘텐츠 추천 효율 향상을 위한 넷플리스 특허 분석 연구

박찬정,김기용 한국차세대컴퓨팅학회 2023 한국차세대컴퓨팅학회 논문지 Vol.19 No.1

With the rapid growth of OTT service providers and content in recent years, there has been a lot of research on techniques to recommend content to users. However, there have been few studies analyzing patents related to content recommendation technologies. In this paper, we analyze patents related to content recommendation among the patents held by Netflix, a leading OTT service, to improve the efficiency of content recommendation. Until June 30, 2022, a total of 431 patents were applied and published by Netflix in the United States, of which 36 patents were related to content recommendation. The 36 patents selected for content recommendation were categorized into rating estimation and recommendation algorithm improvement, content item and genre recommendation, preferred content display and scene display, and other content recommendation technologies. And the patented technologies related to content recommendation were analyzed for each technology area. Based on this, we recommend necessary research in the field of content technology. 최근 OTT 서비스 제공자 및 콘텐츠가 급격히 증가함에 따라, 사용자에게 콘텐츠를 추천하는 기술이 많이 연구되고 있다. 그러나 콘텐츠 추천 기술과 관련된 특허 분석 연구는 거의 이루어지지 않았다. 본 논문에서는 콘텐츠 추천 효율 향상을 위하여 OTT 서비스를 선도하는 넷플릭스가 보유하고 있는 특허 중에서 콘텐츠 추천과 관련된 특허를 분석하였다. 2022년 6월 30일까지 넷플릭스가 미국에 출원하여 공개된 특허는 모두 431건이며, 그 중에서 콘텐츠 추천과 관련된 특허는 36건이었다. 콘텐츠 추천으로 선별된 36건의 특허를 평가 추정 및 추천 알고리즘 개선, 콘텐츠 아이템 및 장르추천, 선호 콘텐츠 디스플레이 및 장면 표시, 기타 콘텐츠 추천 기술로 분류하였고, 각 기술 분야 별로 콘텐츠 추천 관련 특허 기술을 분석하였다. 그리고 이를 토대로 콘텐츠 기술 분야에서 필요한 연구를 제언하였다.

O2O 기반의 상품 추천 및 증강현실 서비스 아키텍쳐 연구

김철진,장태환 한국지식정보기술학회 2022 한국지식정보기술학회 논문지 Vol.17 No.5

온라인 쇼핑몰에서 사용자에게 상품에 대한 구매력을 높이기 위해 추천 기법이 활용되고 있으며 쇼핑의 만족도를 높이기 위해서는 증강현실 서비스나 가상현실 서비스가 활용되고 있다. 본 연구는 오프라인 쇼핑몰에서 온라인 추천 서비스를 제공하고 추천 상품에 대해 증강현실 서비스를 제공하여 오프라인 경험을 제공하기 위한 아키텍처를 제안한다. 추천 서비스는 기존 전통적인 협업적 필터링 기법보다 딥러닝 기법을 활용하여 추천의 정확도를 높일 수 있도록 하고 있다. 본 연구에서는 추천 서비스를 위해 딥러닝 기반의 i-LSTM 알고리즘을 활용하여 추천 정확도가 높은 아키텍처를 제안한다. O2O 기반의 추천 서비스를 제공하기 위해 비콘을 통해 오프라인 상품에 대한 신호를 수신하고 온라인 상품을 추천받는다. 온라인 추천 상품에 대해 쇼핑의 만족도를 향상시키기 위해 오프라인 상품 경험으로 상품에 대한 실재감을 제공하기 위한 증강현실 서비스 아키텍처를 제안한다. 증강현실 서비스는 모바일 어플리케이션과 서버 측면에서 Vuforia 증강현실 개발 플랫폼을 기반으로 한다. 이러한 추천 및 증강현실 서비스 구조를 위한 계층 아키텍처와 흐름 아키텍처를 제안하며 구성 요소와 요소 간의 기능적 흐름을 정의한다. 실험에서는 제안한 O2O 상품 추천 및 증강현실 서비스 아키텍처에 대해 상품 추천 및 증강현실 서비스를 구현하여 아키텍처의 적합성을 검증한다. In an online shopping mall, a recommendation technique is used to increase the purchasing power of a user for a product, and an augmented reality service or a virtual reality service is used to increase shopping satisfaction. This study proposes an architecture for providing an offline experience by providing an online recommendation service in an offline shopping mall and providing an augmented reality service for recommended products. The recommendation service uses deep learning techniques rather than the existing traditional collaborative filtering techniques to increase the accuracy of recommendations. In this study, we propose an architecture with high recommendation accuracy by using deep learning-based i-LSTM algorithm for recommendation service. In order to provide an O2O based recommendation service, signals for offline products are received through beacons and online products are recommended. To improve the satisfaction of shopping for online recommended products, we propose an augmented reality service architecture to provide a sense of reality for products through offline product experiences. The augmented reality service is based on the Vuforia augmented reality development platform in terms of mobile applications and servers. We propose a layer architecture and flow architecture for such recommendation and augmented reality service structures, and define functional flows between components and elements. In the experiment, the suitability of the architecture is verified by implementing the product recommendation and augmented reality service for the proposed O2O product recommendation and augmented reality service architecture.

Assessing Personalized Recommendation Services Using Expectancy Disconfirmation Theory

최일영,문현실,김재경 한국경영정보학회 2019 Asia Pacific Journal of Information Systems Vol.29 No.2

There is an accuracy-diversity dilemma with personalized recommendation services. Some researchers believe that accurate recommendations might reinforce customer satisfaction. However, others claim that highly accurate recommendations and customer satisfaction are not always correlated. Thus, this study attempts to establish the causal factors that determine customer satisfaction with personalized recommendation services to reconcile these incompatible views. This paper employs statistical analyses of simulation to investigate an accuracy-diversity dilemma with personalized recommendation services. To this end, we develop a personalized recommendation system and measured accuracy, diversity, and customer satisfaction using a simulation method. The results show that accurate recommendations positively affected customer satisfaction, whereas diverse recommendations negatively affected customer satisfaction. Also, customer satisfaction was associated with the recommendation product size when neighborhood size was optimal in accuracy. Thus, these results offer insights into personalizing recommendation service providers. The providers must identify customers’ preferences correctly and suggest more accurate recommendations. Furthermore, accuracy is not always improved as the number of product recommendation increases. Accordingly, providers must propose adequate number of product recommendation.

A Content-Based Approach to Recommend TV Programs Enhanced with Delayering Tagging

Fulian Yin,Xingyi Pan,Huixin Liu,Jianping Chai 보안공학연구지원센터 2016 International Journal of Multimedia and Ubiquitous Vol.11 No.9

In response to explore how to extract the recommended items' features, a method is put forward called a Content-based TV Program Recommendation Approach Enhanced with Delayering Tagging. The Content-based approach is optimized to recommend TV programs and improved the way to extract the recommended items' features. Besides, the existing way of using supervised method to build user modeling is replaced with an unsupervised method using delayering tagging to show recommended TV program's content features and set up user preference model. After compared with Latent Factor Model and Collaborative Filtering recommendation algorithm with the same experimental data, the proposed algorithm in this paper increased the accuracy of 2.67\%, coverage rate of 3.02\% and 3.2\% of the Feature 1 value and achieved good recommendation results compared to the Latent Factor Model which revealed the best effect of recommendation.