---
tags: meeting
---

# 12-30 Meeting
## 新版GraphSAGE
Training loss
<!-- 

各自feature normalize -->

<!-- 

 -->

Training MAP
推薦從knn改成dot算score, 且只挑references超過20篇
**要比對一下推薦的部分哪裡有問題**

graph_scores = np.dot(paper_emb_all, paper_emb.T)
graph_recommend_papers = y_all[np.argsort(-graph_scores)][1:K+1]

* MAP@150
<!--  -->

測試一下的Graph Embedding做推薦也有相同問題

目前推測是在算Dot做推薦時有問題，接下來要檢查以下:
- MAP的Answer有沒有錯
- 推薦的部分是不是有問題

<!-- 沒答案的會讓MAP是1.0

 -->

## Optimize Model
## LINE
### First Order and Second Order
### BEACON
* Author: 800/98001
* Correlation Matrix: (16065, 16065)
* Result: