本研究提出一個基於基金持股行為序列的個人投資推薦系統，以解決傳統投資推薦系統面臨的隱私保護和高維度特徵處理等挑戰。現有投資推薦系統多依賴個人交易資料，不僅有可能面臨隱私與安全疑慮，也因高維度特徵而增加建模複雜度。本研究嘗試地採用公開的基金持股資料替代個人資料，透過分析基金經理人的持股行為序列，建立投資組合特徵，並將新投資人的偏好映射至相似的基金投資組合上，間接推薦適合的股票與債券組合。方法上，本研究應用下一籃推薦 (Next Basket Recommendation, NBR) 模型處理基金的時序交易行為，包括基準排序模型、近鄰協同過濾模型及深度學習模型，並比較其在捕捉基金交易行為規律和預測持股變化上的效果。實證結果顯示，與傳統協同過濾和深度學習模型相比，時間感知的近鄰模型 (如Time Influence Factor User K-Nearest Neighbors, TIFUKNN和注意力機制的深度模型 (如Deep Neural Network for Temporal Set Prediction, DNNTSP) 在多數指標上表現最佳。特別是，所有模型在預測重複持股方面表現優異，但在預測新增持股方面表現較弱，反映了基金投資決策的高度穩定性。本研究預期可在規避個人交易資料的隱私問題下，透過專業基金經理的投資智慧，為個人投資者提供更精準的投資建議。;This study proposes a personal investment recommendation system based on fund holdings behavior sequences to address the challenges of privacy protection and high-dimensional feature processing faced by traditional investment recommendation systems. Existing investment recommendation systems largely rely on personal transaction data, which not only faces serious privacy and security risks but also increases modeling complexity due to high-dimensional features. This study innovatively adopts publicly available fund holdings data as an alternative, establishing investment portfolio characteristics by analyzing fund managers′ holdings behavior sequences, and mapping new investors′ preferences to similar fund portfolios to indirectly recommend suitable stock and bond combinations. Methodologically, this study applies Next Basket Recommendation (NBR) models to process funds′ temporal trading behaviors, including baseline ranking models, nearest neighbor collaborative filtering models, and deep learning models, comparing their effectiveness in capturing fund trading behavior patterns and predicting holdings changes. Empirical results show that compared to traditional collaborative filtering and deep learning models, time-aware nearest neighbor models (such as TIFUKNN) and attention mechanism-based deep models (such as DNNTSP) perform best across most metrics. Particularly, all models demonstrate excellent performance in predicting repeated holdings but weaker performance in predicting new holdings, reflecting the high stability of fund investment decisions. This study not only completely circumvents privacy issues associated with personal transaction data but also provides more precise investment advice for individual investors through the investment wisdom of professional fund managers, while outperforming traditional methods in both prediction accuracy and computational efficiency.
