AttGGSNN_model_Overview

Avatar of 黃偉嘉 (Willie Huang).
Avatar of 黃偉嘉 (Willie Huang).

AttGGSNN_model_Overview

Data Scientist / Engineer @ CyberLink
Taipei City, Taiwan
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
Avatar of the user.
Please login to comment.

Published: May 6th 2020
62
8
0

碩士論文

Share