簡易檢索 / 詳目顯示

回結果列表
研究生： 翁品皙
Weng, Pin-Shi
論文名稱： 基於因果時間注意力與多視角跨注意力模型之多模態ESG評等預測
Multimodal ESG Ratings Forecasting using Causality-Aware Temporal Attention and Multi-Perspective Cross-Attention
指導教授： 黃仁暐
Huang, Jen-Wei
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 電機工程學系
Department of Electrical Engineering
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 58
中文關鍵詞： ESG評等 、因果自注意力機制 、跨模態注意力機制 、多模態融合 、機器學習
外文關鍵詞： ESG ratings, Causal self-attention, Cross-modal attention, Multimodal Fusion
相關次數： 點閱：11  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • ESG評等已被廣泛應用於投資決策與企業決策之中。然而，生成ESG評等需花費相對高的成本與時間，主要原因在於評等高度仰賴專家針對多元資料來源進行蒐集、判讀與評估。ESG評等亦常伴隨因財報揭露與報導週期所導致的發布延遲，使得評等較難即時反映企業狀況的快速變化。因此，為及時提供投資人可用的ESG評等資訊，許多研究專注在進行ESG評等預測。然而，現有年度ESG評等預測方法仍存在若干限制。第一，雖然部分研究已指出多模態輸入可提升ESG評等預測效能，但多數方法在模態融合上仍採用較為簡單的策略，例如特徵串接或平均池化，難以充分整合異質模態間的互補資訊。第二，許多既有方法未明確考量年度內關鍵月份的重要性，因而難以捕捉重大事件或揭露對年度評等的影響。為解決上述問題，本研究提出以因果感知的時間注意力捕捉時間相依性並辨識關鍵月份，同時採用多視角交叉注意力機制以進一步提升多模態資訊融合的有效性。實驗結果顯示，我們提出的方法在真實世界ESG資料集上可穩定優於當前最先進的方法，在不同評估設定下最高可改善20.2% 的平均平方誤差（MSE），驗證所提模型之有效性。

    ESG ratings are widely used in investment and corporate decision-making. However, producing ESG ratings is costly and time-consuming because it relies on experts to assess information from diverse sources. Consequently, ratings are typically released with reporting lags, making the ratings slower to reflect rapidly changing company conditions. Therefore, the task of ESG rating forecasting is proposed to provide investors with real time ESG ratings. However, existing approaches for annual ESG rating forecasting remain two limitations. First, although some models have shown that multimodal inputs can improve ESG ratings prediction, they often rely on simple fusion, such as feature concatenation or average pooling to combine heterogeneous modalities. Second, many existing approaches ignore the importance of some key months, hard to capture the influence of material information or disclosure. To address above two challenges, we propose a framework named CATAMCA, which employs Causality-Aware Temporal Attention (CAT) to model historical dependencies and highlight key months. Also, Multi-Perspective Cross-Attention (MCA) is introduced to overcome shallow multimodal fusion and to integrate multimodal information through explicit cross-modal interaction modeling. Experiments on a real-world NYSE dataset show that CATAMCA consistently outperforms strong baselines, achieving up to a 20.2% reduction in MSE and improved IC across evaluation settings, demonstrating the effectiveness of Causality-Aware Temporal and Multi-perspective Cross-Attention.

    中文摘要 ii Abstract iii Acknowledgment iv Table of Contents v List of Tables viii List of Figures ix 1 Introduction 1 2 Related Works 5 2.1 Financial Time-Series Forecasting 5 2.2 Machine Learning Approaches for ESG Ratings Prediction 6 2.2.1 Machine Learning-based Models 6 2.2.2 Deep Learning-based Models 6 2.3 Comparison of Related Works 7 3 The Proposed CATAMCA Framework for ESG Rating Prediction 8 3.1 The Problem Definition of ESG Ratings Prediction 8 3.2 Overview of CATAMCA 8 3.3 Modality-specific Temporal Feature Encoder 10 3.3.1 Sequence Encoder 10 3.3.2 Text Encoder 12 3.4 Causality-Aware Temporal Attention 12 3.4.1 Causal Multi-Head Self-Attention 12 3.4.2 Temporal Attention Aggregator 13 3.5 Multi-Perspective Cross-Attention 14 3.6 ESG Ratings Prediction 15 3.7 Loss Functions 16 4 Experiments 18 4.1 Dataset 18 4.1.1 Stock Prices and Financial Indicators 18 4.1.2 News Articles 18 4.1.3 Event Theme Frequency 21 4.2 Training/Validation/Testing Periods 22 4.3 Comparative Models 22 4.4 Hyperparameter Settings 23 4.5 Evaluation Metrics 23 4.6 Performance Evaluation in NYSE 24 4.6.1 Performance on Predicting the rating of Environment(E) 25 4.6.2 Performance on Predicting the rating of Social(S) 26 4.6.3 Performance on Predicting the rating of Governance(G) 26 4.6.4 Performance on Predicting the rating of Total ESG score 27 4.6.5 Discussion 28 4.7 Ablation Studies 29 4.7.1 Importance of Multi-Perspective Cross-Attention 30 4.7.2 Effectiveness of TAA 31 4.7.3 Analysis of CMHS 31 4.7.4 Assessing the Effectiveness of Individual Input Features 33 4.8 Analysis of Temporal Dynamics and Causal Patterns 34 4.8.1 Qualitative Analysis of Causality-Aware Temporal Attention (E/S/G) 34 4.8.2 Qualitative Interpretation of Temporal Aggregation Weights 36 4.9 Case Study: Identifying Key Months under a Real-World ESG Controversy 41 4.9.1 Causal MHA highlights influential historical months 41 4.9.2 TAA assigns higher weights to controversy-related months, especially from the event modality 41 4.9.3 Prediction Result under year-level supervision 43 5 Conclusions 44 References 46

    [1] Gary Ang, Zhiling Guo, and Ee-Peng Lim. On predicting ESG ratings using dynamic company networks. ACM Transactions on Management Information Systems (TMIS), 14(3):1–34, 2023.
    [2] Gary Ang and Ee-Peng Lim. Learning dynamic multimodal network slot concepts from the web for forecasting environmental, social, and governance ratings. ACM Transactions on the Web, 18(3):1–32, Jun. 2024.
    [3] Tanja Aue, Adam Jatowt, and Michael Färber. Predicting company ESG ratings from news articles using multivariate time-series analysis. In Companion Proceedings of the Web Conference 2025 (WWW ’25 Companion), pages 1774–1780, New York, NY, USA, 2025. ACM.
    [4] Inci M. Baytas, Cao Xiao, Xi Zhang, Fei Wang, Anil K. Jain, and Jiayu Zhou. Patient subtyping via time-aware LSTM networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’17), pages 65–74, New York, NY, USA, 2017. ACM.
    [5] Florian Berg, Julian F. Kölbel, and Roberto Rigobon. Aggregate confusion: The divergence of ESG ratings. Review of Finance, 26(6):1315–1344, 2022.
    [6] Aaron K. Chatterji, Rodolphe Durand, David I. Levine, and Samuel Touboul. Do ratings of firms converge? Implications for managers, investors and strategy. Strategic Management Journal, 37(8):1597–1614, 2016.
    [7] Tianqi Chen and Carlos Guestrin. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 785–794, 2016.
    [8] Edward Choi, Mohammad Taha Bahadori, Jimeng Sun, Joshua Kulas, Walter F. Stewart, and Walter F. Chu. RETAIN: An interpretable predictive model for healthcare using reverse time attention mechanism. In Advances in Neural Information Processing Systems (NeurIPS), volume 29, 2016.
    [9] Jeong-Cheol Choi, Qiuying Chen, and Sang-Joon Lee. Predicting ESG ratings by machine learning and analyzing influencing factors by XAI. In Proceedings of the 2024 7th International Conference on Software Engineering and Information Management (ICSIM ’24), pages 120–125, New York, NY, USA, 2024. Association for Computing Machinery.
    [10] Mohammad Ashraful Ferdous Chowdhury, Mohammad Abdullah, Md. Nazrul Islam, et al. Environmental, social and governance (ESG) rating prediction using machine learning approaches. Annals of Operations Research, pages 1–24, 2023.
    [11] Valeria D’Amato, Rita D’Ecclesia, and Susanna Levantesi. ESG score prediction through random forest algorithm. Computational Management Science, 19(2):347–373, 2022.
    [12] Robert G. Eccles, Ioannis Ioannou, and George Serafeim. The impact of corporate sustainability on organizational processes and performance. Management Science, 60(11):2835–2857, 2014.
    [13] Gunnar Friede, Timo Busch, and Alexander Bassen. ESG and financial performance: Aggregated evidence from more than 2000 empirical studies. Journal of Sustainable Finance & Investment, 5(4):210–233, 2015.
    [14] Tim Krappel, Alex Bogun, and Damian Borth. Heterogeneous ensemble for ESG ratings prediction. arXiv preprint arXiv:2109.10085, 2021. Accepted at KDD Workshop on Machine Learning in Finance (KDD MLF) 2021.
    [15] Hao Liu et al. Temporal heterogeneous graph neural networks for financial time series prediction. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management (CIKM), pages 1038–1047, 2021.
    [16] Ramit Sawhney, Arnav Wadhwa, Shivam Agarwal, and Rajiv Ratn Shah. FAST: Financial news and tweet based time aware network for stock trading. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2253–2261, Online, 2021. Association for Computational Linguistics.
    [17] Tej Bahadur Shahi, Ashish Shrestha, Arjun Neupane, and William Guo. Prediction of ESG rating using machine learning algorithms. Sustainability, 15(7):5910, 2023.
    [18] Andrey Sokolov, Jonathan Mostovoy, Jeff Ding, and Luis Seco. Building machine learning systems for automated ESG scoring. The Journal of Impact and ESG Investing, 1(3):39–50, 2021.
    [19] Yao-Hung Hubert Tsai, Shaojie Shao, Paul Pu Liang, Louis-Philippe Morency, and Ruslan Salakhutdinov. Multimodal transformer for unaligned multimodal language sequences. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL), pages 6558–6569, 2019.
    [20] George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff. A transformer-based framework for multivariate time series representation learning. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (KDD ’21), pages 2114–2124, New York, NY, USA, 2021. ACM.
    [21] Jianfeng Zhang and Zexin Zhao. Corporate ESG rating prediction based on XGBoost-SHAP interpretable machine learning model. Expert Systems with Applications, 295:128809, 2026.

    QR CODE