- 學位論文
政府資訊系統委外案得標特性分析
Analysis on the Characteristics of Winning Bid for Government Information System Outsourcing
摘要
隨著數位化的普及，軟體在我們生活參與的比例越來越高。而數位治理、電子 化政府，已是未來趨勢。但現今的政府採購仍只有對工程類標案，有較完整資源及 監控，其餘勞務、財務等類型的資訊，相對較為零散。在政府採購這個市場中，多 數人都在探討著檯面上的規則，即「政府採購法」。而「潛規測」就是各個提案人 員依自己的投標經驗，一點一點累積而成。也許剛好遇到的競爭對手，都沒有相關 經驗所以獲勝。也許機關太依賴原廠商，而無法放下選擇新廠商。許許多多的情境 也造就了，每個提案人員都有自己的一套，但換一家公司後提案成績就不如以往順 利。也許正是自己提案經驗所建立的「潛規則」並不適用新公司。為更清楚了解資 訊委外標案市場狀況及廠商得標的關鍵點。 本研究以2019年至2023年期間政府採購決標紀錄，關鍵字包含「建置」且履約 內容與系統開發相關之標案。所收集到的標案共有3,432案，進行得標預測，採用 四種分類機器學習方法(羅吉斯迴歸、RBF支持向量機、決策樹及隨機森林)對資 料數據進行分析，在分析過程中，比對未區分資料的準確率及區分後的準確率，另 將機關特徵轉換為需求書特徵，以觀察預測的重要特徵。 結果顯示，資訊委外案的競爭結果是可預測的，在多數情形下，隨機森林有較 高的準確率(77.4%);而「是否為中小企業」、「員工規模」、「公司名稱」、 「投標金額」等特徵都被證實是預測中的重要特徵。建議未來可透過企業形象、專 案實績等因子，對公司名稱所代表的意涵，可進行更深入研究。
並列摘要
With the widespread digitization, software plays an increasingly significant role in our daily lives. Digital governance and e-government have become future trends. However, the current government procurement mainly focuses on engineering projects, with relatively complete resources and monitoring, while information on other types of services and finances tends to be fragmented. In this market, most people are discussing the surface rules, namely the "Government Procurement Law." "Hidden rules" are accumulated by proposal participants based on their bidding experiences. Perhaps competitors who happen to lack relevant experiences win by chance. Perhaps agencies are too reliant on original suppliers and cannot let go to choose new ones. Many scenarios have led to each proposal participant having their own set of rules, but when changing companies, their proposal performance may not be as smooth as before. Perhaps it is because the "hidden rules" established based on one's own proposal experience do not apply to the new company. To gain a clearer understanding of the status and key points of winning bids in the information outsourcing market, this study utilizes government procurement award records from 2019 to 2023, focusing on projects related to "construction" and system development. A total of 3,432 projects were collected for bid prediction. Four classification machine learning methods (logistic regression, RBF support vector machine, decision tree, and random forest) were employed for data analysis. During the analysis, the accuracy rates of undifferentiated data and differentiated data were compared. Additionally, agency features were transformed into demand document features to observe the predictive key features. The results indicate that the outcomes of information outsourcing projects are predictable, with random forest showing higher accuracy rates in most cases (77.4%). Features such as "whether it is a small and medium-sized enterprise," "number of employees," "company name," and "bid amount" have been confirmed as important predictive features. It is suggested that future research can delve deeper into factors such as corporate image and project performance to study the implications represented by company names more comprehensively.