웹 마이닝 기법을 이용한 XML 기반의 개인화된 추천 시스템
(A) Personalized Recommender System based on XML using Web Mining Technique
웹 마이닝 XML 추천 시스템 개인화 기법 PRML;
- 원문 URL
TThis thesis suggests a personalized recommender system model to recommend a specific product or content dynamically, based on a client's taste or preference, by analyzing the individual client's activity pattern employing web mining technique at a XML based e-commerce site. The suggested model is a new technique of personalized recommender system model that can reflect changes in the individual client's preference by studying new case. First, the PRML(Personalized Recommendation Markup Language) technique was suggested to analyze the individual client's activity pattern by using the web mining technique. Next, the content preferred by the user was analyzed by collecting the data on the individual client's activity pattern and saved in the PRML instance. The individual client's profile was also produced by analyzing the preference on the content type accessed by the user by CBR(Case-Based Reasoning) technique. Finally, a new form of personalized recommender system that recommends the upper rated content with a priority by comparing the similarity between the user profile information and the new content was presented. The PRML technique is a technique to collect the personalized identification information such as user ID and IP address effectively by organizing user sessions along with the log information in the CLF (Common Log Format) by session unit. In addition, the technique allows implicit collection of the rating information on the content accessed by the user as well as the attribute information about the XML content type in real time. In this process, the URL accessed by the user, size and the server status are collected and recorded at the PRML instance. The log file size is reduced by removing the images and scripts that are not relevant to the user activity pattern. The rating information about the content accessed by the user carries out an essential role to identify the individual client's preference of the content in the personalized recommender system and to reflect the changes in the individual taste consistently. To collect the individual's preference of the content effectively, the technique that collects the rating information implicitly using the hierarchical structure of the XML content and the element weight was suggested. The rating information calculates the individual client's rating information recorded at the PRML instance implicitly by using the weight database by element and the element set of the content accessed by the user. This value means the individual preference about the content accessed by the user. Additionally, a technique that collects the attribute information using CBR technique to retrieve the individual preference about the XML content type was suggested. The intra-attribute weight and the inter-attribute weight were also defined in order to express the individual preference as the attribute information about the content type. The individual user profile was also produced from the CBR attribute information recorded at the PRML instance. The user profile means the preference about the content type that the user accessed, and it is also used by the CBR technique for the purpose of recommending the new content. The recommendation technique suggested in this thesis is unlike the collaborative filtering technique using the rating value of other users, one of the techniques to recommend the content that the user prefers by analyzing the individual activity pattern, and it is the model that is most similar to the CBR technique widely used in the area of AI. The new content for each individual is recommended by the process in which the similarity of the preferences between the new content and the each content type saved in the user profile is calculated and the new content is recommended by the order of similarity. An experiment to recommend the new content for the user of the content site was conducted to prove the effectiveness of the personalized recommender system model suggested in this thesis. First, a comparison was made with the existing CLF form by collecting the user identification information by the PRML technique. The suggested technique showed that the personalized log information can be effectively retrieved, which is difficult to collect in the CLF form. The experiment to collect the implicit rating information showed that the user preference can be more effectively retrieved by this technique than by the existing explicit rating information technique that causes inconvenience to the user or the existing implicit rating information collection technique that mostly uses the content reading time. The individual profile was organized to analyze the preference about the content type that the user accessed in order to conduct the content recommendation research. Then, the experiment to recommend the new content for each individual was conducted by comparing the similarity between the user profile and the new content. A comparison was made on the recommendation efficiency among the collaborative filtering technique used most widely by MAE and ROC measurement value, the demographic filtering technique, the content-based technique and the technique suggested in this study. The result of the experiment showed that the technique that this study suggests shows the highest recommendation efficiency. This model can be applied to the web mining technique to develop a personalized service by analyzing the user's activity pattern at a XLM based e-commerce site. In addition, it is also expected to be applicable to the eCRM technique to maintain a close relationship with the clients in the area of e-commerce by analyzing the client's taste and to provide a differentiated service to the preferred clients.