DEVELOPMENT OF DTLPLP FOR LINK PREDICTION ON DYNAMIC HETEROGENEOUS NETWORK USING HETER-LP APPROACH

• المؤلف الرئيسي: Hemsa Ghaida, Radhinansyah
• التنسيق: Theses
• اللغة: Indonesia
• الوصول للمادة أونلاين: https://digilib.itb.ac.id/gdl/view/77863
• المؤسسة: Institut Teknologi Bandung

Link prediction is a method to see the possibility of a link in a complex network based on the previously formed network. However, until now the link prediction method has only been developed for heterogeneous or dynamic cases. In fact, most real cases can be described with heterogeneous and dynamic networks. The implementation of this thesis is carried out to carry out the development of the DTPLLP method which can perform link prediction in dynamic homogeneous cases with the addition of the Heter- LP concept which performs link prediction in static heterogeneous cases so that the proposed method can perform link prediction in dynamic heterogeneous cases. The main concept added to the proposed changes to the DTLPLP algorithm is the bipartite network projection concept so that the proposed algorithm takes into account the heterogeneous network characteristics. In addition, the dynamic characteristics of the network are considered while maintaining the compression stage of the network. Then, changes were made to the link prediction stage based on label propagation so that link prediction can be carried out on heterogeneous networks. The test was carried out using the Yelp dataset on April 1-7 2020 with the AUC-ROC evaluation metric. Based on the tests carried out, the best scores for predicting all links were 0,4796; 0,6392; 0,9787 respectively for user-business, user-user, and user- business link types. In addition, the best scores for predicting new links were 0,9486; 0,7595; 0,8787 for user-business, user-user, and user-business link types, respectively. The development of the DTLPLP method by adding the heterogeneous concept of Heter-LP has been successfully carried out. It can be concluded that the proposed algorithm is good for predicting new links that appear but not good for predicting all links consisting of new links and lost links.