2017-06-012024-05-18https://scholars.lib.ntu.edu.tw/handle/123456789/702159摘要：針對網路商務未來發展，此整合型計畫提出一個「整合異質性資料之智能推薦架構」，有效結 合文字、圖片、音訊、社群、使用者行為等異質性資料，適用於各種不同應用情境，達成智 能推薦之目標。本計畫將分四項子計畫執行︰中研院楊奕軒教授將針對異質性資料開發基於 矩陣分解之推薦演算法；政大蔡銘峰教授將專注於探討如何運用網路表示式學習法於推薦問 題；中研院王釧茹教授將專注於利用離散的文字資訊學習適切之物品及使用者表示式；臺大 陳宏銘教授則將利用深度學習開發連續型資料（例如︰音樂、影像訊號）之表示式學習法。 此整合型計畫的四項子計畫，各有其負責重點，但彼此關聯且互相支援。除了學術貢獻，本 整合型計畫也將透過下列方式，對我國網路電子商務產業的研發與成長產生實質助益：1、開 放研究結果之相關工具（如：基於圖形表示式學習法之推薦工具模組化函式庫）；2、公開所 蒐集整理之資料集（如：社群評論、社群網路關係之資料）；3、提供相關學習處理過後的資 源（如：特定領域之自然語言處理相關資源、經機器學習並優化後的物件表示向量）；4、開 放相關應用情境之雛型（如：個人化音樂推薦、菜單推薦系統雛型）。<br> Abstract: We propose a three-year joint project to develop a computational framework for processing and understanding heterogeneous data (including texts, pictures, audio signals, social relations, and user behaviors) and using the embedding representations thus learned to design intelligent recommendation schemes for various e-commerce services. The project is divided into four sub-projects. Prof. Yi-Hsuan Yang from Academia Sinica is responsible for the development of recommendation algorithms based on matrix factorization; Prof. Ming-Feng Tsai from National Chengchi University is responsible for the development of network embedding learning algorithms for recommendation; Prof. Chuan-Ju Wang from Academia Sinica is responsible for the development of item and user concept representation learning from discrete heterogeneous data; Prof. Homer Chen from National Taiwan University is responsible for the development of representation learning for continuous heterogeneous data. The four sub-projects are closely interlinked and operated in a collaborative manner to ensure a successful delivery of the following materials at the end of the three-year research period: 1) open-source library of graph embedding learning techniques and associated recommendation algorithms, 2) datasets such as community reviews and relations of social network data collected during the course of the joint project, 3) relevant learned data resources including natural language processing related resources and machine optimized vector representations of objects, and 4) application prototypes such as personalized music recommendation and menu recommendation.異質性資料矩陣分解表示法學習深度學習自然語言處理數值訊號處理Heterogeneous datamatrix factorizationdata embeddingdeep learningnatural language processingnumerical signal processing