http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Kotaro Yamashita,Ayako Yamagiwa,Kyosuke Hasumoto,Masayuki Goto 대한산업공학회 2024 Industrial Engineeering & Management Systems Vol.23 No.2
In recent years, many product reviews have been posted on e-commerce sites. These review data contain the impressions and requests of users who have purchased and used the products and have a direct impact on the purchasing behavior of other users. On the other hand, manufacturers need to analyze review data not only to understand users’ needs but also to understand the problems of existing products. In addition, since these review data include the evaluation values of the users who submitted the data, the analysis of the review data is valuable as direct product evaluation information by the users. Although users are generally expected to give a rating that matches the content of the review, some users are dissatisfied with the product but give a high rating. Conversely, there are some users who are satisfied with the product but give an intermediate rating in a cursory way, so the content of the review and the rating do not necessarily match. In such cases, judging the evaluation of a product by focusing on the evaluation value may result in a product evaluation that deviates from the actual evaluation by the user. Therefore, this study introduces BERT (Bidirectional Encoder Representations from Transformers) and sentiment analysis methods, which have recently shown effectiveness in natural language processing. We propose a method for estimating evaluation values that consider the users emotional content expressed in the review sentences. Furthermore, we apply the proposed method to the real review data and demonstrate its usefulness.
Yuka Nakamura,Taiga Yoshikawa,Ayako Yamagiwa,Masayuki Goto 대한산업공학회 2024 Industrial Engineeering & Management Systems Vol.23 No.2
In recent years, many recommender systems are introduced in e-commerce sites, which estimate each users preferences based on the past log data and show a list of his preferable items. The performance of these systems is usually evaluated by measuring whether presented recommendation lists match the preferences of customers. Here, the recommender system does not only need to provide a list just once for each user, but also it is in fact a continuous recommending process. It is, therefore, important to discuss the performance based on the cumulative loss for series of recommendation. Meanwhile, online learning is a framework that can handle such problem of sequential ecommendation and evaluation. However, the purpose of online learning is to improve the efficiency of learning in order to estimate condition of best values, and most methods do not consider the cumulative loss. On the other hand, Safe Exploration for Optimization (SafeOpt) was proposed as a method to perform exploration while suppressing the deterioration of the objective function not only the best value. This method has, however, a problem that the searchable range depends on initial values. In this study, this paper extends the original SafeOpt by introducing GP-UCB and proposes a method to enable a wide search area while suppressing the deterioration of the objective function. To show the effectiveness of the proposed method, the simulation experiments by using artificial data are shown.