• Proceedings of the Korea Information Processing Society Conference
• /
• 2017.04a
• /
• pp.439-441
• /
• 2017

Matrix factorization은 사용자의 아이템 선호도를 통해 아이템을 추천해주는 성공적인 기술 중 하나이다. 이 기법은 사용자-아이템의 선호도 행렬을 채우는 것을 목표로 한다. 이 목표를 달성하기 위해 사용자-아이템의 선호도 행렬을 사용자 행렬(user latent factor)와 아이템 행렬(item latent factor)로 분해하고, 각 행렬에 대해 추론하여 완성된 사용자-아이템의 선호도 행렬을 추론한다. 하지만 Matrix factorization은 아이템의 수가 많고, 아이템에 대한 사용자들의 선호도 데이터가 적을 때 성능이 제한된다. 또한 새로운 아이템이 추가되었을 때, 새로운 아이템에 대한 사용자들의 선호도 정보가 없기 때문에 새로운 아이템이 추천되지 않는다는 문제를 가진다. 이를 해결하기 위해 본 논문에서는 아이템에 대한 부가적인 정보인 아이템 간의 유사도 정보와 아이템의 시나리오 정보의 유사도를 모델링하여 기존의 전통적인 Matrix factorization에 추가하는 아이템 정보 기반 추천 시스템을 제안한다.

• Journal of Intelligence and Information Systems
• /
• v.28 no.3
• /
• pp.119-137
• /
• 2022

Recent recommendation system studies apply various deep learning models to represent user and item interactions better. One of the noteworthy studies is ONCF(Outer product-based Neural Collaborative Filtering) which builds a two-dimensional interaction map via outer product and employs CNN (Convolutional Neural Networks) to learn high-order correlations from the map. However, ONCF has limitations in recommendation performance due to the problems with CNN and the absence of side information. ONCF using CNN has an inductive bias problem that causes poor performances for data with a distribution that does not appear in the training data. This paper proposes to employ a Vision Transformer (ViT) instead of the vanilla CNN used in ONCF. The reason is that ViT showed better results than state-of-the-art CNN in many image classification cases. In addition, we propose a new architecture to reflect side information that ONCF did not consider. Unlike previous studies that reflect side information in a neural network using simple input combination methods, this study uses an independent auxiliary classifier to reflect side information more effectively in the recommender system. ONCF used a single latent vector for user and item, but in this study, a channel is constructed using multiple vectors to enable the model to learn more diverse expressions and to obtain an ensemble effect. The experiments showed our deep learning model improved performance in recommendation compared to ONCF.

• Information Systems Review
• /
• v.23 no.1
• /
• pp.155-172
• /
• 2021

With the recent increase in online product purchases, a recommender system that recommends products considering users' preferences has still been studied. The recommender system provides personalized product recommendation services to users. Collaborative Filtering (CF) using user ratings on products is one of the most widely used recommendation algorithms. During CF, the item-based method identifies the user's product by using ratings left on the product purchased by the user and obtains the similarity between the purchased product and the unpurchased product. CF takes a lot of time to calculate the similarity between products. In particular, it takes more time when using text-based big data such as review data of Amazon store. This paper suggests a hybrid recommendation system using a 2-phase methodology and text data mining to calculate the similarity between products easily and quickly. To this end, we collected about 980,000 online consumer ratings and review data from the online commerce store, Amazon Kinder Store. As a result of several experiments, it was confirmed that the suggested hybrid recommendation system reflecting the user's rating and review data has resulted in similar recommendation time, but higher accuracy compared to the CF-based benchmark recommender systems. Therefore, the suggested system is expected to increase the user's satisfaction and increase its sales.

• Journal of the Institute of Electronics Engineers of Korea CI
• /
• v.43 no.4 s.310
• /
• pp.50-57
• /
• 2006

In Electronic Commerce, the latest most of the personalized recommender systems have applied to the collaborative filtering technique. This method calculates the weight of similarity among users who have a similar preference degree in order to predict and recommend the item which hits to propensity of users. In this case, we commonly use Pearson Correlation Coefficient. However, this method is feasible to calculate a correlation if only there are the items that two users evaluated a preference degree in common. Accordingly, the accuracy of prediction falls. The weight of similarity can affect not only the case which predicts the item which hits to propensity of users, but also the performance of the personalized recommender system. In this study, we verify the improvement of the prediction accuracy through an experiment after observing the rule of the weight of similarity applying Vector similarity, Entropy, Inverse user frequency, and Default voting of Information Retrieval field. The result shows that the method combining the weight of similarity using the Entropy with Default voting got the most efficient performance.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.21 no.5
• /
• pp.221-226
• /
• 2021

Collaborative filtering based recommender systems recommend user-preferrable items based on rating history and are essential function for the current various commercial purposes. In order to determine items to recommend, prediction of preference score for unrated items is estimated based on similar rating history. Previous studies usually employ two methods individually, i.e., similar user based or similar item based ones. These methods have drawbacks of degrading prediction accuracy in case of sparse user ratings data or when having difficulty with finding similar users or items. This study suggests a new rating prediction method by integrating the two previous methods. The proposed method has the advantage of consulting more similar ratings, thus improving the recommendation quality. The experimental results reveal that our method significantly improve the performance of previous methods, in terms of prediction accuracy, relevance level of recommended items, and that of recommended item ranks with a sparse dataset. With a rather dense dataset, it outperforms the previous methods in terms of prediction accuracy and shows comparable results in other metrics.

• The Journal of Korean Institute of Information Technology
• /
• v.16 no.12
• /
• pp.109-114
• /
• 2018

Collaborative filtering has been widely utilized in recommender systems as typical algorithm for outstanding performance. Since it depends on item rating history structurally, The more sparse rating matrix is, the lower its recommendation accuracy is, and sometimes it is totally useless. Variety of hybrid approaches have tried to combine collaborative filtering and content-based method for improving the sparsity issue in rating matrix. In this study, a new method is suggested for the same purpose, but with different perspective, it deals with no-match situation in person-person similarity evaluation. This method is called the transitive similarity model because it is based on relation graph of people, and it compares recommendation accuracy by applying to Movielens open dataset.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.15 no.4
• /
• pp.877-885
• /
• 2011

A collaborative filtering(CF) is the most widely used technique in recommender system. However, CF has sparsity and scalability problems. These problems reduce the accuracy of recommendation and extensive studies have been made to solve these problems, In this paper, we proposed a method that uses a weight so as to solve these problems. After creating a user-item matrix, the proposed method analyzes information about users who prefer the item only by using data with a rating over 4 for enhancing the accuracy in the recommendation. The proposed method uses information about the genre of the item as well as analyzed user information as a weight during the calculation of similarity, and it calculates prediction by using only data for which the similarity is over a threshold and uses the data as the rating value of unrated data. It is possible simultaneously to reduce sparsity and to improve accuracy by calculating prediction through an analysis of the characteristics of an item. Also, it is possible to conduct a quick classification based on the analyzed information once a new item and a user are registered. The experiment result indicated that the proposed method has been more enhanced the accuracy, compared to item based, genre based methods.

• Asia pacific journal of information systems
• /
• v.20 no.1
• /
• pp.81-99
• /
• 2010

Lately, in consumers' markets the number of new items is rapidly increasing at an overwhelming rate while consumers have limited access to information about those new products in making a sensible, well-informed purchase. Therefore, item providers and customers need a system which recommends right items to right customers. Also, whenever new items are released, for instance, the recommender system specializing in new items can help item providers locate and identify potential customers. Currently, new items are being added to an existing system without being specially noted to consumers, making it difficult for consumers to identify and evaluate new products introduced in the markets. Most of previous approaches for recommender systems have to rely on the usage history of customers. For new items, this content-based (CB) approach is simply not available for the system to recommend those new items to potential consumers. Although collaborative filtering (CF) approach is not directly applicable to solve the new item problem, it would be a good idea to use the basic principle of CF which identifies similar customers, i,e. neighbors, and recommend items to those customers who have liked the similar items in the past. This research aims to suggest a hybrid recommendation procedure based on the preference boundary of target customer. We suggest the hybrid recommendation procedure using the preference boundary in the feature space for recommending new items only. The basic principle is that if a new item belongs within the preference boundary of a target customer, then it is evaluated to be preferred by the customer. Customers' preferences and characteristics of items including new items are represented in a feature space, and the scope or boundary of the target customer's preference is extended to those of neighbors'. The new item recommendation procedure consists of three steps. The first step is analyzing the profile of items, which are represented as k-dimensional feature values. The second step is to determine the representative point of the target customer's preference boundary, the centroid, based on a personal information set. To determine the centroid of preference boundary of a target customer, three algorithms are developed in this research: one is using the centroid of a target customer only (TC), the other is using centroid of a (dummy) big target customer that is composed of a target customer and his/her neighbors (BC), and another is using centroids of a target customer and his/her neighbors (NC). The third step is to determine the range of the preference boundary, the radius. The suggested algorithm Is using the average distance (AD) between the centroid and all purchased items. We test whether the CF-based approach to determine the centroid of the preference boundary improves the recommendation quality or not. For this purpose, we develop two hybrid algorithms, BC and NC, which use neighbors when deciding centroid of the preference boundary. To test the validity of hybrid algorithms, BC and NC, we developed CB-algorithm, TC, which uses target customers only. We measured effectiveness scores of suggested algorithms and compared them through a series of experiments with a set of real mobile image transaction data. We spilt the period between 1st June 2004 and 31st July and the period between 1st August and 31st August 2004 as a training set and a test set, respectively. The training set Is used to make the preference boundary, and the test set is used to evaluate the performance of the suggested hybrid recommendation procedure. The main aim of this research Is to compare the hybrid recommendation algorithm with the CB algorithm. To evaluate the performance of each algorithm, we compare the purchased new item list in test period with the recommended item list which is recommended by suggested algorithms. So we employ the evaluation metric to hit the ratio for evaluating our algorithms. The hit ratio is defined as the ratio of the hit set size to the recommended set size. The hit set size means the number of success of recommendations in our experiment, and the test set size means the number of purchased items during the test period. Experimental test result shows the hit ratio of BC and NC is bigger than that of TC. This means using neighbors Is more effective to recommend new items. That is hybrid algorithm using CF is more effective when recommending to consumers new items than the algorithm using only CB. The reason of the smaller hit ratio of BC than that of NC is that BC is defined as a dummy or virtual customer who purchased all items of target customers' and neighbors'. That is centroid of BC often shifts from that of TC, so it tends to reflect skewed characters of target customer. So the recommendation algorithm using NC shows the best hit ratio, because NC has sufficient information about target customers and their neighbors without damaging the information about the target customers.

• Journal of KIISE:Software and Applications
• /
• v.32 no.7
• /
• pp.581-591
• /
• 2005

Collaborative filtering systems have problems involving sparsity and the provision of recommendations by making correlations between only two users' preferences. These systems recommend items based only on the preferences without taking in to account the contents of the items. As a result, the accuracy of recommendations depends on the data from user-rated items. When users rate items, it can be expected that not all users ran do so earnestly. This brings down the accuracy of recommendations. This paper proposes a collaborative recommendation method for extracting typical group preferences using user-item matrix optimization and user profiles in collaborative tittering systems. The method excludes unproven users by using entropy based on data from user-rated items and groups users into clusters after generating user profiles, and then extracts typical group preferences. The proposed method generates collaborative user profiles by using association word mining to reflect contents as well as preferences of items and groups users into clusters based on the profiles by using the vector space model and the K-means algorithm. To compensate for the shortcoming of providing recommendations using correlations between only two user preferences, the proposed method extracts typical preferences of groups using the entropy theory The typical preferences are extracted by combining user entropies with item preferences. The recommender system using typical group preferences solves the problem caused by recommendations based on preferences rated incorrectly by users and reduces time for retrieving the most similar users in groups.

• Journal of Intelligence and Information Systems
• /
• v.29 no.3
• /
• pp.167-183
• /
• 2023

In the recent field of recommendation systems, various studies have been conducted to model sparse data effectively. Among these, GLocal-K(Global and Local Kernels for Recommender Systems) is a research endeavor combining global and local kernels to provide personalized recommendations by considering global data patterns and individual user characteristics. However, due to its utilization of kernel tricks, GLocal-K exhibits diminished performance on highly sparse data and struggles to offer recommendations for new users or items due to the absence of side information. In this paper, to address these limitations of GLocal-K, we propose the GEase-K (Global and EASE kernels for Recommender Systems) model, incorporating the EASE(Embarrassingly Shallow Autoencoders for Sparse Data) model and leveraging side information. Initially, we substitute EASE for the local kernel in GLocal-K to enhance recommendation performance on highly sparse data. EASE, functioning as a simple linear operational structure, is an autoencoder that performs highly on extremely sparse data through regularization and learning item similarity. Additionally, we utilize side information to alleviate the cold-start problem. We enhance the understanding of user-item similarities by employing a conditional autoencoder structure during the training process to incorporate side information. In conclusion, GEase-K demonstrates resilience in highly sparse data and cold-start situations by combining linear and nonlinear structures and utilizing side information. Experimental results show that GEase-K outperforms GLocal-K based on the RMSE and MAE metrics on the highly sparse GoodReads and ModCloth datasets. Furthermore, in cold-start experiments divided into four groups using the GoodReads and ModCloth datasets, GEase-K denotes superior performance compared to GLocal-K.