---
tags: 暑假
---

# 0728
## 建文
### Temporal類的做法
- SOTA
- 跑baseline
- 提主要idea

**Related work:**
> Keywords: dynamic graph, temporal
> [推薦常見的datasets](https://github.com/DeepGraphLearning/RecommenderSystems/blob/master/dataList.md)

1. Session-based Social Recommendation via Dynamic Graph Attention Networks ==Dynamic Graph, Graphsage==
2. Déjà vu: A Contextualized Temporal Attention Mechanism for Sequential Recommendation
3. [Graph contextualized self-attention network for session-based recommendation](https://www.ijcai.org/proceedings/2019/0547.pdf) ==Dynamic Graph==
4. [Next-item Recommendation with Sequential Hypergraphs]() ==Hypergraph==
---

:::spoiler
5. [Deep Spatio-Temporal Neural Networks for Click-Through Rate Prediction](https://arxiv.org/pdf/1906.03776.pdf) ==CTR==
6. [Hierarchical Temporal Convolutional Networks for Dynamic Recommender Systems](https://arxiv.org/pdf/1904.04381.pdf)
7. [A Dynamic Co-attention Network for Session-based Recommendation](https://staff.fnwi.uva.nl/m.derijke/wp-content/papercite-data/pdf/chen-2019-dynamic.pdf)
8. [Online Purchase Prediction via Multi-Scale Modeling of Behavior Dynamics](https://xuczhang.github.io/papers/kdd19_purchase_pred.pdf)

---
:::

### SOTAs

**Long & short-term interest**
- NARM
- STAMP
- SR-GNN
- TA-GNN

**可能可以發展的方向:**
- 同item在不同時間意義會不同
- 同user/ session在不同時刻的preference不同
- 同items組成的session, 會因順序不同而有不同意義
- sliding window/ attention => 抓不同區段對next action的影響

:::spoiler
### MANN(Memory-Augmented Neural Networks)
當我們train大型的Nets時, 如果可以把過程中的parameters值儲存起來, 當遇到類似或新的資料進來時靠有效的檢索系統把那些值即時讀出來, 會比多次重複性質的Gradient Descent會有效的多. 所以架構簡單的來說就是有個個controller控制input-output(其實有點類似海馬迴或電腦的CPU), 然後有Memory區(相等於大腦中的新皮層或電腦的RAM). 而利用另一個類神經網路去控制controller及讀取頭去精確控制要寫入或讀取的記憶區塊.


### Aggregate

:::

## TY
Graph類models
> https://airtable.com/shrH8vpKlF8U0stSz

### SOTAs
- SR-GNN
- TA-GNN
- A-PGNN

**可以當baseline的其他方法**
- NARM
- STAMP
- GRU4Rec
- Caser
- SASRec