Search

6. Knowledge Graph Based Recommendation

Person
๋น„๊ณ 

6.0 Knowledge Graph(KG) intro

โ€ข
Social network : user representation ์ž˜ ๋ฝ‘๊ธฐ ์œ„ํ•ด users relationships ๋ฐ˜์˜
โ€ข
KG : item representation ์ž˜ ๋ฝ‘๊ธฐ ์œ„ํ•ด items(attribute-์†์„ฑ) relationships ํ‘œํ˜„
โ€ข
KG๋ฅผ ์ถ”์ฒœ์— ํ™œ์šฉ ์‹œ ์žฅ์ 
1.
์•„์ดํ…œ ์‚ฌ์ด์˜ ํ’๋ถ€ํ•œ ์˜๋ฏธ๋ก ์  ๊ด€๊ณ„๋ฅผ ํƒ์ƒ‰, ํ™œ์šฉ ๊ฐ€๋Šฅ
2.
์œ ์ €์˜ ๊ณผ๊ฑฐ ์‚ฌ์šฉ ์ƒํ’ˆ(์ƒํ’ˆ๊ณผ์˜ ์ƒํ˜ธ ์ž‘์šฉ), ์ถ”์ฒœ ํ•ญ๋ชฉ ์ •๋ณด๋ฅผ ์ด์šฉํ•ด ์ถ”์ฒœ ๊ฒฐ๊ณผ์˜ ํ•ด์„๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ž„
โ€ข
ํ•˜์ง€๋งŒ KG์˜ ๋ณต์žกํ•œ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ ๋•Œ๋ฌธ์— ํ™œ์šฉ ์‰ฝ์ง€ ์•Š์Œ
โ—ฆ
์ด์ „ ์—ฐ๊ตฌ : knowlege grapph embedding (KGE) method๋กœ KG ์ „์ฒ˜๋ฆฌ โ†’ link prediction์— ํšจ๊ณผ์ ์ด์ง€๋งŒ ์ถ”์ฒœ์—๋Š” ๋ถ€์ ํ•ฉ
โ€ข
KG ์™ธ์—๋„ user-item interaction ์ •๋ณด๊นŒ์ง€ ์ฃผ์–ด์กŒ์„ ๋•Œ ๋ณด๋‹ค ํšจ๊ณผ์ 
โ—ฆ
์œ ์ € ์„ ํ˜ธ๋„ ์˜ˆ์ธก์— ํ•„์š”ํ•œ item ์‚ฌ์ด์˜ ๊ด€๊ณ„์„ฑ explicitly capture ๊ฐ€๋Šฅ
Reference : RippleNet
โ€ข
KGR์˜ ์ฃผ์š” ์ด์Šˆ
1.
Graph Construction
a.
user-item ์ƒํ˜ธ์ž‘์šฉ์— ๋‚ดํฌ๋œ collaborative signal๊ณผ ์™€ KG์˜ ์˜๋ฏธ์ ์ธ ์ •๋ณด๋ฅผ ์–ด๋–ป๊ฒŒ ํšจ๊ณผ์ ์œผ๋กœ ํ†ตํ•ฉ?
b.
์œ ์ € ๋…ธ๋“œ๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ KG์— ํ†ตํ•ฉํ•  ๊ฒƒ์ด์ง€, ํ˜น์€ ๊ด€๊ณ„์˜ ์ค‘์š”์„ฑ์„ ๊ตฌ๋ณ„ํ•˜๊ธฐ ์œ„ํ•ด ์œ ์ € ๋…ธ๋“œ๋ฅผ implicitํ•˜๊ฒŒ ์‚ฌ์šฉ?
2.
Relation-aware Aggregation
a.
KG์˜ ํ•œ ํŠน์ง•์€ entities ์‚ฌ์ด์— ๊ด€๊ณ„๊ฐ€ ์—ฌ๋Ÿฌ type์œผ๋กœ ์ด๋ฃจ์–ด์ง
b.
์ด๋Ÿฌํ•œ ์—ฐ๊ฒฐ๋œ entity๋“ค ์‚ฌ์ด์˜ ์ •๋ณด๋ฅผ ํ†ตํ•ฉํ•˜๊ธฐ ์œ„ํ•ด relation-aware aggregation function์„ ์–ด๋–ป๊ฒŒ ๋””์ž์ธ?

6.1 Graph Construction

โ€ข
How to effectively integrate the collaborative signals and knowledge information?
1.
user node๋ฅผ KG์— ํ†ตํ•ฉ (unified graph)
Reference : Hierarchical Attentive Knowledge Graph Embedding for Personalized Recommendation
โ€ข
unified graph๋ฅผ ๊ตฌ์„ฑํ•˜๋Š” ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ - KGAT, MKGAT, CKAN, AKGE โ€ฆ
2.
user node๋ฅผ implicit ํ•˜๊ฒŒ ์‚ฌ์šฉ - ์„œ๋กœ ๋‹ค๋ฅธ ๊ด€๊ณ„๋“ค์˜ ์ค‘์š”๋„ ๊ตฌ๋ณ„ / ex) KGCN, KGNN-LS
Reference : KGNN-LS
โ€ข
user-item subgraph ๋ฐฉ์‹์€ ๊ด€๋ จ์žˆ๋Š” entity์™€ ๊ด€๊ณ„์— ๋Œ€ํ•ด ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ์ง€๋งŒ, computation time ๋” ๋งŽ์ด ํ•„์š”ํ•˜๊ณ  subgraph ๊ตฌ์„ฑ์— ๋”ฐ๋ผ ์„ฑ๋Šฅ์ด depend ๋˜์–ด ์•„์ง ๋” ๋งŽ์€ ์—ฐ๊ตฌ ํ•„์š”

6.2 Relation-aware Aggregation

โ€ข
KG์˜ ์˜๋ฏธ์ ์ธ ์ •๋ณด๋ฅผ ์ œ๋Œ€๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” linked entities, relations ๋ชจ๋‘ propagation process์— ๊ณ ๋ ค๋˜์–ด์•ผ ํ•จ
โ€ข
ex) KGAT : relation์— weight ํ• ๋‹น
โ€ข
WrW_r : transformation matrix for the relation, which maps the entity into relation space
โ€ข
the closer entities would pass more information to the central node
โ€ข
user node๋ฅผ ํฌํ•จํ•˜๋Š” subgraph์— ์ ์šฉํ•˜๋Š”๊ฒŒ ๋” ์ ์ ˆ

6.3 Summary