請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/52376
標題: | 基於異質資訊網路表示法學習之孤兒保單業務員推薦 Orphan Policy’s Agent Recommendation Based on Heterogeneous Information Network Embedding |
作者: | Pei-Yu Ko 柯沛瑜 |
指導教授: | 曹承礎(Seng-Cho Chou) |
關鍵字: | 人壽保險,業務員,孤兒保單,網路表示法,推薦系統, life insurance,agents,orphan policy,network embedding,recommendation system, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 在保險業中,當保單的業務員無法繼續服務該保單,該保單就會變成「孤兒保單」,而保險公司就必須替這些孤兒保單指派新的業務員。在臺灣,人壽保險業務員的流動率大,讓顧客保單容易變成孤兒保單,讓保險公司也需常進行業務員的調撥,然而目前指派的業務員並沒有明確的依據與規定,而當新業務員不適合顧客時,就可能造成保險公司顧客流失。因此本篇研究將提出一個顧客與業務員間的推薦系統,來協助保險公司找到適合顧客的好業務員。 本研究透過業務員與顧客的購買關係網路,透過網路表示法(Network Embedding),將顧客與業務員之節點投影於低為度的向量空間中,並基於此向量空間建立推薦系統,推薦顧客未來會與之購買的業務員,並以Hit Ratio@K來評估模型的表現。其中,本研究使用兩種不同的網路表示法學習節點向量,分別為沒有用節點特徵的node2vec與納入節點特徵的GraphSAGE,並比較兩種網路表示方法的表現差異。 實驗結果顯示,未納入節點特徵的node2vec會有較佳的推薦表現。本研究提出的推薦系統,對有經過調撥顧客未來購買保單業務員的Hit ratio@10可以達到0.63。而對於顧客跟非調撥業務員購買的情況,被視為不當調撥的情況,也有不錯的推薦表現,代表本研究的推薦模型可以對那些不當調撥的顧客,在調撥當時就找到顧客未來會與之購買保單的業務員。 In the insurance industry, when an insurance policy agent cannot continue to serve the policy, the policy will become “orphan policy”, and the insurance company must reallocate agents to these orphan policies. In Taiwan, the turnover rate of new insurance agents is high, which makes customer insurance policies easier to become orphan policies, and insurance companies also need to frequently reallocate agents. However, the current agents reallocated have no clear basis and regulations. When a reassigned agent is not suitable for customers, it may cause insurance companies to lose their customers. Therefore, this study will propose a recommendation system between customers and agents to help companies find good agents suitable for customers. In this study, we establish a purchase relationship network between agents and customers. Through network embedding, the nodes of customers and salespersons in the purchased network are projected into a low-dimension vector space, and a recommendation system is established based on this vector space. And we use Hit Ratio@K to evaluate the performance of the model. This study uses two network embedding algorithm to learn node vectors, namely node2vec and GraphSAGE, and compares the performance differences between the two methods. Experimental results show that node2vec will have better recommendation performance. The recommendation system proposed in this study can reach 0.63 for the Hit ratio@10. As for the situation with the customer who buy the policies with non-reallocation of sales, is considered improper reallocation of the case, also has a good performance. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/52376 |
DOI: | 10.6342/NTU202002582 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0608202021332900.pdf 目前未授權公開取用 | 2.8 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。