| 摘要: | 企業融資限制不僅會降低其未來的發展潛力，更進一步會弱化整個產業的競爭力，故許多學 者相繼提出企業融資限制的評估指標，但過去的評估指標僅利用統計方法捕捉到資料間的線 性結構，並無法捕捉到資料間的非線性結構，為了克服此問題，本計畫將拆解支援向量機的 內部結構並將其與急速學習機做整合，以捕捉資料間的非線性結構，為了更進一步提升模型 的預測效力，將利用主題分析技術將大量文字型態的年報資訊做拆解並彙整成數個指標，以 補充數值型態無法提供的資訊。此外，利用社會網絡分析技術探討企業的商業網絡關係，並 評估此企業在此商業網絡的影響力，更進一步將此關係拆解為多層面利害關係人觀點予以探 討，即企業供應鏈網絡關係、企業競爭者供應鏈網絡、協同合作者供應鏈網絡，並分析此關 係之影響力，最後，將此混合模型的內部決策邏輯予以呈現，對於知識萃取的技術將採用學 習型、拆解型與次經驗演算法等，並考慮認知適配理論與知識品質的評估指標，而同時考量 多個層面會導致多元組合的問題，有鑑於此，本計劃將此問題轉為多目標決策分析的問題， 並利用多目標決策分析的技術以系統性的方式予以解決，並提出一個易於使使用者瞭解的企 業融資限制的決策輔助預測模型。
Development of the financial sector has considerable influence on society that facilitates the economic development and technological innovation. In well-structured economies, financial intermediaries help to finance tangible and intangible investments, thereby enabling corporates to pursue advanced development. However, not all of the financial market are well-functioning that will cause some problem (i.e., financial constraints) impedes the corporate development. Thus, numerous researchers laid much more emphasis to construct the assessment criterions to evaluate the corporate financial constraints. However, most criterions are based on statistical-based technique, such as multivariate discriminant analysis (MDA) that can merely capture the data with linear structure. That is, MDA is not suitable to handle the data with non-linear structure. To overcome this obstacle, this study proposes a hybrid model that can be categorized into four parts: (1) textual information extraction, (2) business relation definition, (3) forecasting model construction, and (4) knowledge representation. Textual information were gathered from corporate’s annual reports. By executing text mining technique and topic modelling, the considerable textual information can be condensed into some manageable ones. Corporate’s business relation is another important competitive advantage in today’s highly fluctuated economic market. How to evaluate the importance of corporate embedded in business network is a critical problem. This study collects the business information from annual report and database and implements social network analysis technique to construct the corporate business network. By implementing social network technique, we can realize the importance of the corporate embedded in the business network. This study further divided the corporate business relation into three stakeholder aspects: supply chain relation, competitor relation, and cooperator relation, and examine each impact on corporate financial constraints. Sequentially, the numerical information, textual information, business relation information was fed into support vector-based extreme learning machine (SV-ELM) to construct the forecasting model. The SV-ELM not only preserves the advantages of SVM (i.e., superior generalizability, outstanding forecasting accuracy), but also sustains the merits of ELM (i.e., efficiency). This study implements SV-ELM to capture the non-linear structure of the data. One of the critical drawbacks of hybrid model is lacking of comprehensibility. To overcome this challenge, this study used three knowledge extraction techniques: pedagogical, de-compositional, and meta heuristic, to represent the decision logics embedded into hybrid model. In order to increase the acceptance by decision-makers, the knowledge representation styles also take into consideration. It is widely recognized that dissimilar knowledge representation styles will lead to dissimilar acceptance rate by decision-makers. This problem can be transformed into multiple criteria decision analysis (MCDA) task and MCDA algorithm (called VIKOR) can solve it. The proposed corporate financial constraints forecasting model can be taken as decision assisted support system by decision-makers to form their own judgments. |
