# 0814 Meeting
## Motivation
- SR-GNN等graph based model沒有考慮到
1. aggregate後的pattern順序性
2. aggregate時應該考慮幾層鄰居的資訊
- PA-GNN對node embedding加上reverse order embedding來給予attention一些順序上的資訊

- **目標: 把所有node embeeding塞進有考慮時間序列的model**
- 類似的pattern會在不同時間上出現
- 2017: iphone X -> airpods 1 -> iphone X cases
2019: iphone 11 -> airpods pro -> iphone 11 cases

## Problem Settings
- 同個item在不同時間點的session被點擊時應該會具有不同的物理意義
- 考慮不同action的重要性 -> view & click

## Baselines
- A Contextualized Temporal Attention Mechanism for Sequential Recommendation ==缺前處理code==
> https://tianchi.aliyun.com/dataset/dataDetail?dataId=649



- ==**Make It a Chorus: Knowledge- and Time-aware Item Modeling for Sequential Recommendation**==



- Time Interval Aware Self-Attention for Sequential Recommendation
- 對position和time interval做embedding
- 換成session資料 => jdata優先, 因為時間範圍也很廣



- Time Matters: Sequential Recommendation with Complex Temporal Information
- Absolute Time Pattern
- customers are more likely to buy T-shirts in summer than in winter
- ==Relative Time Pattern==
- focuses on the **time interval between each pair of user behaviors**
- The green line presents a user’s repeated purchases of guitar picks
- guitar->tuner (red line): Most users purchase guitars and tuner together
- guitar->strings (blue line): buy new strings about 150 days later than the guitar


- Global Context Enhanced Graph Neural Networks for Session-based Recommendation


:::spoiler
:::

---

:::spoiler
- DySAT: Deep Neural Representation Learning on Dynamic Graphs via Self-Attention Networks

---

- GACOforRec: Session-Based Graph Convolutional Neural Networks Recommendation Model

- Modeling Item-Specific Temporal Dynamics of Repeat Consumption for Recommender Systems

:::

---

## Experiments
> https://docs.google.com/spreadsheets/d/1zt0VoYEde4HbgbkTIJVProU2dTSEA0xDXtnyDM3aRA4/edit?usp=sharing


- jdata, Tabaoo, Tmall, yochoose 1/4待補
- kkbox沒有timestamp

> yoochoose 1/4 跑1 epoch大該需train 7小時
> 

其他篇論文給的數據

---
某區user
A區關注iphone8 -> 換B區關注 ->換C區
隔年, A區關注iphone11 -> 換B區關注 ->換C區

- 上方改成dynamic graph (放入user/ session id), 讓不同時間的item有不同embedding, 最後和原本的session embedding結合