# 0518-Meeting

<!-- ## Last-Click為相同item的比例
 -->

## Rosetta.ai
<!-- ### config.transformed_dummy_item

### continuous_features
 -->


<!--  -->

**藍色區塊:**

**橘色區塊:Previous interacted**

**黃色區塊:過去互動的type, click等**

<!-- 過去互動過的item
 -->

**綠色區塊:**
這個session內扣掉要預測的item後的impression

Continuous Feature

V dot U

Max-pooling 原始做法：

Click-out items(model最左feature)
取出一個max過後的item, action embedding

## Concat All
切割檔案會導致新mapping的id跑掉

## Process time
train 25w sessions, val 5w

| Process | Time |
| -------- | -------- |
| Preprocess | 1249s |
| Train | 813s |
| Val | 47s |

---
train 50w sessions, val 5w

| Process | Time |
| -------- | -------- |
| Preprocess | 2542s |
| Train | 1646s |
| Val | 58s |

---

## Attentive RNN


6隻前處理code共會產生34隻csv
### Pandas vs Pickle

Pandas存太慢 (30分鐘才存2G)

trade-off


單一個dataframe, 在session只用30%的情況還會到30G+

### Performance
10W session (10%), MRR 0.65
Claim 0.678


<!-- | Process | Time |
| -------- | -------- |
| Preprocess | 大於4hr |
| Train | 1hr |
| Val | > 5 mins | -->

### 專題生進度安排
> https://share.clickup.com/c/h/3fjyn-159/c4e47a282ec9173