利用融資訊息和圖神經網路進行新創創投鏈接推薦

Abstract

新創企業是社會中重要的創新來源，卻受困於極高的失敗率。超過三成早期倒閉的新創將失敗的主因歸咎於募資困難。創業投資家（VC）作為新創企業最重要的資金來源之一，從合適的VC取得融資對於新創的成功有很大幫助。然而，新創企業與創投之間存在著高度資訊不對稱，導致配對困難。本研究旨在開發一種基於深度學習的推薦系統，協助新創企業找到適合的投資者。 本研究的核心貢獻在於將融資活動的時間序列資料引入新創與 VC 的關係建模中，並以新創視角將過往一對一的鏈結預測問題視為 VC 排序問題，致力於推薦合適的投資者給新創公司。我們取用 SDC VentureXpert 資料庫 2009 年至 2020 年期間的新創企業融資資料。我們提出了一個深度學習模型，結合了靜態公司特徵、動態融資資訊以及動態 VC 共投關係，以預測新創企業與 VC 之間的投資配對可能性。 為了評估模型效能，我們採用了全類平均準確率（mAP）、平均倒數排名（MRR）和前 K 準確率（Precision@K）等指標。實驗結果顯示，我們的模型在所有指標上均優於傳統模型，證明了深度學習與時間序列資料在解決新創與 VC 匹配問題上的有效性。此外，消融實驗結果顯示，公司特徵以及過往投資者相關特徵對於提升模型效能至關重要。 本研究的發現對於新創企業和 VC 具有重要的實務意涵，通過精準的推薦系統，新創企業可以更有效地尋找到潛在投資者，VC 也能接觸更多具有潛力的投資機會。然而，本研究仍存在一些限制，例如未提供第一輪預測、資料集的完整性等，未來研究可以進一步擴大研究範圍，並整合不同資料源的資訊。

Startups are a critical engine of innovation but face high failure rates. Over 30% of early-stage startups attribute their failure to difficulties in securing funding. Venture capitalists (VCs) are a primary source of funding for startups, and securing investment from suitable VCs is crucial for their success. However, the information asymmetry between startups and VCs hinders effective matching. This study aims to develop a deep learning-based recommendation system to assist startups in finding suitable investors. The core contribution of this study lies in incorporating sequential data of funding activities into the modeling of relationships between startups and VCs. By adopting a startup-centric perspective, we transform the traditional one-to-one link prediction problem into a VC ranking problem, focusing on recommending the most suitable investors for each startup. Using funding data from the SDC VentureXpert database spanning 2009 to 2020, we propose a deep learning model that combines startup profile features, funding-related features, historical investor behaviors, round information, and dynamic VC co-investment relationships to predict the likelihood of investment between startups and VCs. To evaluate the model’s performance, we employ metrics such as mean average precision (mAP), mean reciprocal rank (MRR), and precision@K. Experimental results show that our model outperforms traditional models on all metrics, demonstrating the effectiveness of deep learning and sequential data in solving the startup-VC matching problem. Moreover, ablation studies reveal that company features and historical investor-related features are crucial for improving model performance. The findings of this study have significant practical implications for both startups and VCs. By providing an accurate recommendation system, startups can more efficiently identify potential investors, while VCs can discover more promising investment opportunities. However, the study has some limitations, such as the lack of first-round predictions and data completeness. Future research can further expand the scope and integrate information from various data sources.