書目資料為中介的協力過濾式推薦於圖書館之應用

Abstract

推薦系統（Recommendation System）是圖書館協助使用者篩選資料的一種工具，但因多數資料缺乏使用紀錄，使得推薦系統面臨資料稀疏等問題，影響推薦效能。本研究以臺北市立圖書館的書目資料和2013至2018年的借閱紀錄為基礎，選擇「題名」、「作者」、「分類號」、「主題標目」、「出版社」等5種書目資料，提出以書目資料作為中介之混和推薦方法，檢視其預測效果。 本研究以2013年至2017年的借閱紀錄為訓練資料集、2018年借閱紀錄為答案資料集，將訓練資料集的書籍借閱紀錄轉換為書目資料借閱紀錄，再依據書目資料的共被借閱率以Message Passing Clustering（MPC）分群方法進行階層分群，建構5個單種書目資料的相似度矩陣。利用相似度矩陣將個人借閱紀錄轉為興趣特徵，得出預測結果。最後與答案資料集中同一人之借閱紀錄相比較，透過預測分數、預測結果餘弦相似度與前N名預測結果評估5種書目各自與整合的預測結果。 評估結果顯示「主題標目」與「出版社」的整體預測分數最佳，單次借閱均分大於5.5，但是前20名預測結果中，「題名」、「作者」成功預測答案資料集的次數較多，分別有2萬5,939次與6,841次。考量真實情境下使用者的認知負荷，且「題名」、「作者」與「出版社」預測結果之排序成高度相關，建議選擇預測成功次數較多的「題名」取代「作者」，並將「出版社」的預測結果作為補充。此外，「主題標目」預測結果與其他4種書目資料呈低度相關，亦可提供使用者不同面向的新資訊。本研究針對書目資料用於混和推薦系統之結果可提供未來書籍推薦方法研究以及實務運用之參考。

Recommendation system is a tool that libraries can assist users to filter data. Lack of usage records for most data leads to challenges, such as data sparsity. This study proposes a hybrid recommendation method based on the Taipei Public Library borrowing records from 2013 to 2018. Five types of bibliographic data: title, author, classification number, subject heading, and publisher, were selected as intermediaries to deal with data sparsity issues. This study transformed borrowing records to bibliographic data borrowing records. Then, it hierarchically clustered these bibliographic data by co-borrowed rates with the Message Passing Clustering (MPC) method. The clustered results were utilized to construct 5 similarity matrices for each type of bibliographic data. Personal records were converted into interest features by these matrices and obtain predictions. Finally, the predictions of each type of bibliographic data and the integrated predictions were evaluated based on prediction score, cosine similarity, and the top N predictions. The results show that "subject heading" and "publisher" have better prediction scores. Average scores of records are above 5.5. However, within the top 20 predictions, "title" and "author" are better, with 25,939 and 6,841 success predictions, respectively. Considering user''s cognitive load and high correlation between "title," "author," and "publisher," this study suggest to use "title," which has more success predictions, rather than "author." "Publisher" can be used as a supplementary source. Additionally, since the predictions of "subject heading" are less correlated with the other four types of bibliographic data, it can provide users some different perspectives of new information. The results of this study on using bibliographic data in hybrid recommendation system can serve as a reference for future research and practical applications in book recommendation methods.