//assets.cakeresume.com/assets/contentarea/contentarea-2d64eca8dae2a1ffb31e0fb1df887bbad072c7c19734d7d30fa14884bbcc299a.css

我的碩士論文：AttGGSNN所構成的問答系統，採用類似VQA的架構，以過去一段時間的所有DJIA30公司股價return作為輸入(V)，配合“格式限制好的“([公司A, 公司B, 問題種類])問題(Q)，輸出答案(A)

VQA-like system

Data Scientist / Engineer @ CyberLink

黃偉嘉 (Willie Huang)

我的碩士論文：AttGGSNN的模型總覽，主要分成三個module - Feature embedding, Graph propagator, and Question embedding
功能簡述：
1. Feature embedding：將return data先透過1DCNN, GRU等NN做初步的feature extraction，輸出的feature map配合相鄰矩陣(多個類別的edge、edge介於0-1之間)，以作為Graph Propagator的初始狀態
2. Graph propagator：一個大module，主要分成Graph Aggregator與Graph Updater，前者用來蒐集相鄰節點資訊，後者用來更新Graph的node state
2-1. Graph Aggregator：實作Multi-head的Graph attention (包含node attention與edge attention)，先做node attention (self attention)，再將不同edge的結果做edge attention (dot-product attention)，最後輸出Aggregated message (經過attention mechanism所蒐集的相鄰節點資訊)
2-2. Graph Updater：實作一個GRU-like updater，將Aggregated message以及上一個時間點的node state作為輸入，透過GRU的門控機制更新node state
3. Question embedding：將特化的Question以multi-hot或distributed representation vector表示
最後將Final step的node state與3.產生的latent code concatenate起來，過一層DNN並做multi-output的multi-class classification

AttGGSNN_model_Overview

I have a Master's degree in Information Management with a focus on deep Learning and Fintech and currently working as a data scientist and engineer at CyberLink.
I have built 3 machine learning related data projects / 1 data products from scratch.
I am committed to helping companies build stable, efficient, and accurate machine learning systems with the spirit of MLOps.

【學歷經歷】
經歷：
・Data Scientist / Engineer @ CyberLink
(2020.06 - Now)
    - Data Pipeline (Airflow)
    - Data Visualization (Superset)
    - Data Warehouse (Kylin)
    - MLOps (MLflow, SageMaker)
    - Machine / Deep Learning
    - Python / SQL / Gremlin
    - AWS Services
    - Git / Linux / Docker
學歷：
・國立台南一中
・國立成功大學工業與資訊管理系學士
・國立交通大學資訊管理研究所碩士
論文發表：
頂級期刊第一作者
Huang, W. C., Chen, C. T., Lee, C., Kuo, F. H., & Huang, S. H. (2023). Attentive gated graph sequence neural network-based time-series information fusion for financial trading. Information Fusion, 91, 261-276.

【比賽獲獎】
「第22屆全國大專校院資訊應用服務創新競賽」
・兩岸交流組　第二名
・資訊技術應用組　佳作

【產學合作】
・工研院 -「投資標的資料探勘技術」
以python、深度學習方法，進行多種金融商品間交互關係研究及建模。
・成大財金所 -「Form 8-K重大事件財報文字探勘」
以python、網路爬蟲、統計、機器學習等方法，去分析美國各大上市公司發布的重大消息對其股價的影響。

【碩士論文】
「基於改良式GGSNN與注意力機制發展財務時間序列關聯問答應用」
透過深度學習方法「圖神經網路」建構一個商品之間的交互關係模型，並藉由注意力機制聚焦於具高度交互關係的商品上，最終透過此模型預測這些商品間的未來市場趨勢。

1. NCTU
Teaching Assistant of "Advanced Programming (2018)

2. Master's Thesis
"A Question Answering System for Financial Time-Series Correlation Based on Improved Gated Graph Sequence Neural Network with Attention Mechanism"	08/2018 - 08/2019 
- An attention-based Graph Neural Network for Financial Data
- A Deep Learning based Robo-Advisor prototype, to design a QA system for the purpose of recommendation of profolio
- Modeling the correlation between time-series sequences or market information
- About 2.5 accumulated profit in the experiment (DJIA 30, 2017)

3. ITRI
Industry-Academia Cooperation (2018.03 - 2018.12)

國立交通大學

國立成功大學

Cyberlink

Data Scientist

• Focus on applying data science, machine learning, and business intelligence to work 
• Build data pipeline for requirements via Apache Airflow
• Data ETL
    - Deal with > 10 million records/GB level of daily server logs and market data 
• AWS Services
    - S3, Athena, SageMaker, Neptune, Kinesis, Personalize, Rekognition, and so on 
• Data Visualization/Analysis/Report
    - Apache Superset, Tableau, Google Analytics
    - ad-hoc request with Jinja template 
• Data Warehouse, OLAP - Apache Kylin 
• Machine Learning
    - MLOps - MLflow
    - SOTA paper survey, implementation and deployment
    - Domain: Recommendation system, Graph neural network, Computer vision
• Build 3 data products from scratch 
    (1) Content retrieval system
        - Deep learning model to extract content information from templates, automatically tagging to template
        - Huge improvement of template exploration and user experience after conducting statistic test (Improved by over 30%)
    (2) Personalized recommendation system
        - Launch on Promeo - App launched by CyberLink
        - A from-scratch MLOps based recommendation system that automatically and steadily train, choose, and deploy model then inference
        - Pass real-world performance test via A/B test that have a conversion rate enhancement by over 35%
    (3) Graph neural network based Customer Data system
        - Construct GNN-based Persona and conduct customer segmentation 
        - Model users and their interaction and relationship into graph structure 
        - Explore the implicit preference of customers with the help of GNN

分享