Conflict of Interest based Features for Expert Classification in Bibliographic Network

• Published in: 2018 International Conference on Computer Engineering, Network and Intelligent Multimedia (CENIM) pp. 54 - 59
• Main Authors: Purwitasari, Diana, Ilmi, Akhmad Bakhrul, Fatichah, Chastine, Fauzi, Willy Achmat, Sumpeno, Surya, Purnomo, Mauridhi Hery
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2018
• Subjects: bibliographic data; Citation analysis; Collaboration; conflict of interest feature; Deep learning; expert classification; Feature extraction; Google; Metadata; word embedding
• DOI: 10.1109/CENIM.2018.8710931

Countless approaches of feature extraction in the expert classification problem employ text contents and network structures from bibliographic metadata of published articles. The content part often use title and abstract while the structure part utilize co-authorship and citation. On citation data, the classifier method works on a feature of citation quantity since a frequently cited author is presumed to have more expertise. Citation misconduct occurs if there is no subject relation between citing and cited articles. Therefore, the misconduct becomes a challenge for evaluation of citation quality. Here, the problem is to classify experts with features that can indicate citation misconduct. To address this problem, our contribution exploited the quality and the quantity of citations in feature extraction designed for classifying experts. Co-authorship that influence the misconducts is called as Conflict of Interest (CoI) situation. Accordingly, the class labels are experts with or without CoI indication. We proposed three ratio features of (1) self-citation to represent the citation quantity, then (2) subject similarity of author interests and article contents, as well as (3) subject similarity of citing and cited articles to determine the citation quality. There are various word phrases used in subjects with similar contexts. Therefore the proposed CoI-based features for the citation quality took on deep learning approaches for understanding natural language. Our experiments exercised a selection of data from one of the common datasets in bibliographic related problems called as AMiner. We selected ± 15K articles from the original data of ± 2M articles in the experiments. The results showed that our proposed features classified experts with CoI indication by accuracy value of ± 60%. Although the first feature of citation quantity was not significant for categorizing experts, other features of citation quality confirmed more profound evidence.