Model for Session-based Recommendation ADVISOR: JIA-LING, KOH - - PowerPoint PPT Presentation

STAMP Short-Term Attention Memory Priority Model for Session-based Recommendation

ADVISOR: JIA-LING, KOH SOURCE: KDD 2018 SPEAKER: SHAO-WEI, HUANG DATE: 2019/07/01

1

⚫ Introduction ⚫ Method ⚫ Experiment ⚫ Conclusion

2

OUTLINE

INTRODUCTION

➢ What is a session?

3

A user open a broswer. A user close a broswer.

A session Long term memory Short term memory (Last click)

INTRODUCTION

➢ What is session-based recommendation?

⚫ Predict user’s next action (click on an item) in

the current session. 4 Input：user’s clicks in a session

Predict user’s next click item

Output：a socre (the probability

• f an item)

5

INTRODUCTION

STAMP Short-Term Attention/Memory Priority

➢ Only consider Long term memory, without explicitly taking

into account that users’ interests drift with time.

Defects of traditional session-based recommendation

OUTLINE

Introduction Method Experiment Conclusion

6

METHOD

7

Short-Term Memory priority(STMP)

⚫ 𝑛𝑡 denotes the user’s general interest. ⚫ 𝑛𝑢 denotes the user’s current interest.

➢Input： 𝑛𝑢 = 𝑦𝑢

• 𝑦𝑗：𝑓𝑛𝑐𝑓𝑒𝑒𝑗𝑜𝑕 𝑤𝑓𝑑𝑢𝑝𝑠𝑡 𝑝𝑔 𝑢ℎ𝑓 𝑗𝑢𝑓𝑛 𝑗

(d-dimensional)

A session

METHOD

8

Short-Term Memory priority(STMP)

⚫ A simple MLP without hidden layer

is used for feature abstraction.

• 𝑔 · 𝑗𝑡 𝑏 𝑜𝑝𝑜 − 𝑚𝑗𝑜𝑓𝑏𝑠

𝑏𝑑𝑢𝑗𝑤𝑏𝑢𝑗𝑝𝑜 𝑔𝑣𝑜𝑑𝑢𝑗𝑝𝑜 (𝑢𝑏𝑜ℎ)

➢MLP layer：

⚫ ℎ𝑡 = 𝑔 𝑋

𝑡 𝑛𝑡 + 𝑐𝑡

ℎ𝑢 = 𝑔 𝑋

𝑢 𝑛𝑢 + 𝑐𝑢

d*1 d*d d*1 d*1 d*d d*1

candidate

METHOD

9

Short-Term Memory priority(STMP)

⚫ Calculate the trilinear product of the

candidate i. ➢Predict score：

candidate

<ℎ𝑡 ,ℎ𝑢 ,𝑦𝑗 > = σ𝑙=1

𝑒

ℎ𝑡𝑙 ℎ𝑢𝑙 𝑦𝑗𝑙

⚫ Use softmax function to obtain

probability distribution of items.

• 𝑦𝑗：𝑓𝑛𝑐𝑓𝑒𝑒𝑗𝑜𝑕 𝑤𝑓𝑑𝑢𝑝𝑠𝑡 𝑝𝑔

𝑢ℎ𝑓 𝑗𝑢𝑓𝑛 𝑗(d-dimensional)

• σ：sigmoid function

METHOD

10

Short-Term Memory priority(STMP)

⚫ Cross entropy

➢Loss function：

candidate

• |V|：𝐷𝑏𝑜𝑒𝑗𝑏𝑒𝑏𝑢𝑓 item dictionary

the ground truth the ground truth

METHOD

11

Short-Term Attention Memory priority(STAMP)

⚫ Generate attent weights

➢Add attention net into STMP：

candidate

• 𝑦𝑗：𝑓𝑛𝑐𝑓𝑒𝑒𝑗𝑜𝑕 𝑤𝑓𝑑𝑢𝑝𝑠𝑡 𝑝𝑔

𝑢ℎ𝑓 𝑗𝑢𝑓𝑛 𝑗(d-dimensional)

• σ：sigmoid function

𝑦𝑗

⚫ Compose attent weights

OUTLINE

Introduction Method Experiment Conclusion

12

13

EXPERIMENT

Dataset

➢ Yoochoose:

• Consists of six months of click-streams gathered from an e-

commerce web.

• 7,966,257 sessions of 31,637,239 clicks on 37,483 items.

➢ Diginetna:

• Also consists click-streams gathered from an e-commerce

web.

• 202,633 sessions of 982,961 clicks on 43,097 items.

14

EXPERIMENT

➢ Evaluation metrics

• N ：測次資料的總比數
• 𝑜ℎ𝑗𝑢 ：正確答案有被排在預測出來ranking list前K名的筆數。
• Rank(t) ：正確答案被排在ranking list的第幾名。

15

EXPERIMENT

➢ Next-click prediction on 3 benchmark data sets

Neural model

Not neural model

16

EXPERIMENT

➢ Compare STAMP with NARM

17

EXPERIMENT

➢ Effects of the last click

• −：𝑜𝑝𝑢 𝑣𝑡𝑓 𝑚𝑏𝑡𝑢 𝑑𝑚𝑗𝑑𝑙 𝑗𝑢𝑓𝑛 𝑓𝑛𝑐𝑓𝑒𝑒𝑗𝑜𝑕

𝑗𝑜 𝑢ℎ𝑓 𝑢𝑠𝑗𝑚𝑗𝑜𝑓𝑏𝑠 𝑚𝑏𝑧𝑓𝑠.

18

EXPERIMENT

➢ Effects of the last click

• n Yoochoose 1/64

19

EXPERIMENT

➢ Effects of the last click

• n Yoochoose 1/64

20

EXPERIMENT

➢ Comparison among proposed models

• n Yoochoose
• n Diginetica
• Short：sessions length <=5.
• Long：sessions length >5.

Short is better than Long Long is better than Short

21

EXPERIMENT

➢ Comparison among proposed models

• on Yoochoose → Short is better than Long
• on Diginetica → Long is better than short

Short Long

**Repeated clicks ratio have an inversely proportional ratio to model performance

22

EXPERIMENT

➢ Further Investigation（Attention）

Session ID Categort of an ID

OUTLINE

Introduction Method Experiment Conclusion

23

24

CONCLUSION

➢ Propose a short-term attention/memory priority model for session-

based recommendations.

➢ The next move of a user is mostly affected by the last-click of a

session prefix, and our model can effectively utilize such information through the temporal interests representation.

➢ The proposed attention mechanism can effectively capture long-

term and short-term interests of a session.