Search

3. User-Item Collaborative Filtering

Person
๋น„๊ณ 
index
โ€ข
CF๋Š” ์œ ์ €์™€ ์•„์ดํ…œ์˜ ์ƒํ˜ธ์ž‘์šฉ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ, ์œ ์ €์—๊ฒŒ ์„ ํƒ๋œ ์  ์žˆ๋Š” ์•„์ดํ…œ๋“ค์˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•ด user representation์„ ๊ตฌ์„ฑํ•˜๊ณ , ์•„์ดํ…œ๊ณผ ์ƒํ˜ธ์ž‘์šฉํ•œ ์œ ์ €๋“ค์˜ ์ •๋ณด๋ฅผ ์‚ฌ์šฉํ•ด item representation์„ ๊ตฌ์„ฑํ•˜๋Š”(์„œ๋กœ ์ •๋ณด๋ฅผ ์ฃผ๊ณ ๋ฐ›๋Š”) ํ˜•ํƒœ
โ€ข
GNN์€ ๋…ธ๋“œ(user, item) ๊ฐ„ high-order connectivity(interactions)์—์„œ์˜ ์ •๋ณด ํ™•์‚ฐ์„ ์ž˜ ๋ชจ๋ธ๋งํ•  ์ˆ˜ ์žˆ์–ด CF์— ์œ ์šฉ
โ€ข
(๊ธฐ์กด์˜ CF๊ฐ€ user, item representation์„ ๋‹จ์ˆœ ํ•™์Šต ๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋กœ ๋’€๋‹ค๋ฉด, GNN์ด ์ ์šฉ๋œ CF์—์„œ๋Š” user, item representation์„ user-item interaction bipartite graph ์ƒ์—์„œ GNN์œผ๋กœ embedding propagation ์ง„ํ–‰ํ•˜์—ฌ final representation์„ ์ƒ์„ฑํ•จ)

0. 4 Issues

1. Graph Construction

โ€ข
Heterogeneous(user-item) bipartite graphe ์ƒ์—์„œ GNN ์ ์šฉํ• ๊นŒ? ์•„๋‹ˆ๋ฉด user graph / item graph ๊ฐ๊ฐ ๋งŒ๋“ค์–ด์„œ GNN ์ ์šฉํ• ๊นŒ?
โ€ข
Full graph ์ƒ์—์„œ GNN ์ ์šฉํ• ๊นŒ? ๋น„์‹ธ๋‹ˆ๊นŒ ์ด์›ƒ๋“ค์„ ์ƒ˜ํ”Œ๋งํ•ด์„œ GNN ์ ์šฉํ• ๊นŒ?

2. Neighbor Aggregation

โ€ข
์ด์›ƒ๋“ค์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋งค๊ฒจ์„œ ์ •๋ณด Aggregateํ•˜๋ฉด ์–ด๋–จ๊นŒ?
โ€ข
์ด์›ƒ๋“ค ๊ฐ„์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ๋ฐ˜์˜ํ•˜๋ฉด ์–ด๋–จ๊นŒ?

3. Information Update

โ€ข
์ค‘์‹ฌ ๋…ธ๋“œ์™€ ์ฃผ๋ณ€ ๋…ธ๋“œ ์ •๋ณด๋ฅผ ์–ด๋–ป๊ฒŒ ํ•ฉ์น  ๊ฒƒ์ธ๊ฐ€?

4. Final Node Representation

โ€ข
final representation์œผ๋กœ ๋งˆ์ง€๋ง‰ layer rep๋งŒ ์‚ฌ์šฉํ•  ๊ฒƒ์ธ๊ฐ€? ๋ชจ๋“  layer rep์„ ๋‹ค ์‚ฌ์šฉํ•  ๊ฒƒ์ธ๊ฐ€?

1. Graph Construction

1-1. Bipartite Graph๋งŒ ์‚ฌ์šฉํ•  ๊ฒƒ์ธ๊ฐ€

Two-Hop Edges: BIpartite graph์— two-hop edges ๋”ํ•˜๋ฉด user-user / item-item subgraph๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ. ์ด๋ฅผ ํ†ตํ•ด user-user๊ฐ„, item-item๊ฐ„ ์ •๋ณด ์ „๋‹ฌ์ด ํ™œ๋ฐœํžˆ ๋ฐœ์ƒ
Virtual Nodes
โ€ข
DGCF: Bipartite + Two-Hop Edges + ๊ฐ interaction์— ๋Œ€ํ•œ Intent๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” Virtual Nodes๋ฅผ ์ถ”๊ฐ€
โ€ข
Multi-GCCF: Bipartite + Two-Hop Edges
โ€ข
DHCF: HyperEdge ์‚ฌ์šฉํ•ด high-order correlation ๋ชจ๋ธ๋ง
โ€ข
HiGNN: ๋น„์Šทํ•œ ์œ ์ € / ์•„์ดํ…œ์„ ํด๋Ÿฌ์Šคํ„ฐ๋งํ•˜๊ณ  ํด๋Ÿฌ์Šคํ„ฐ ์ค‘์‹ฌ์„ Virtual Nodes๋กœ ์žก์•„ ์ƒˆ๋กœ์šด coarsened user-item graph๋ฅผ ๊ตฌ์„ฑ: user, item๋“ค์˜ ๊ณ„์ธต์  ์—ฐ๊ด€์„ฑ์„ ๋ชจ๋ธ๋ง

1-2. ์ด์›ƒ ์ƒ˜ํ”Œ๋ง

โ€ข
Multi-GCCF, NIA-GCN: ๊ณ ์ • ๊ฐฏ์ˆ˜ ๋žœ๋ค ์ƒ˜ํ”Œ๋ง
โ€ข
Pinsage: ์ค‘์‹ฌ ๋…ธ๋“œ์—์„œ random walk ๋Œ๋ ค์„œ ์ฃผ๋ณ€ ๋…ธ๋“œ๋“ค์—์˜ visit count๋ฅผ ์„ผ ํ›„, ์ƒ์œ„ t๊ฐœ ์ด์›ƒ์„ ์„ ํƒ

2. Neighbor Aggregation

2-a. Equal Weight

โ€ข
๋‹จ์ˆœ Mean-Pooling

2-b. Degree Normalization

โ€ข
GCN์—์„œ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•

2-c. Attentive Weights

โ€ข
์ค‘์‹ฌ ๋…ธ๋“œ์™€ ๋” ๊ด€๋ จ ์žˆ๋Š” ์ฃผ๋ณ€ ๋…ธ๋“œ์ผ์ˆ˜๋ก ๊ฐ€์ค‘ํ•˜๊ณ  ์‹ถ๋‹ค: attention mechanism ์‚ฌ์šฉ
โ€ข
MCCF

2-d. Central Node Augmentation

โ€ข
h_i, h_u ์˜ elementwise product vector์„ ์‚ฌ์šฉ โ†’ h_i, h_u๊ฐ€ align์ด ์ž˜ ๋งž๋Š”์ง€๋ฅผ ๋ฐ˜์˜ํ•œ๋‹ค.
โ€ข

2-e. Neighbor Interaction

โ€ข
์ด์›ƒ๋“ค ๊ฐ„์˜ interaction์„ aggregation ๊ณผ์ •์—์„œ ๋ฐ˜์˜
โ€ข
NIA-GCN

3. Information Update

3-a. ์ฃผ๋ณ€ ๋…ธ๋“œ์˜ ์ •๋ณด๋งŒ์„ ์‚ฌ์šฉํ•ด ์ƒˆ๋กœ์šด Representation ์ƒ์„ฑ

h: ์ค‘์‹ฌ ๋…ธ๋“œ ์ •๋ณด / n: ์ฃผ๋ณ€ ๋…ธ๋“œ ์ •๋ณด
ex)
with transformation/nonlinearity
sum/avg
LR-GCCF: without nonlinearity
LightGCN: without transformation/nonlinearity

3-b. ์ค‘์‹ฌ ๋…ธ๋“œ์™€ ์ฃผ๋ณ€ ๋…ธ๋“œ ์ •๋ณด๋ฅผ ๋ชจ๋‘ ์‚ฌ์šฉ

ex)
with concat / transformation / nonlinearity

4. Final Node Representation

โ€ข
mainstream: ๋งˆ์ง€๋ง‰ layer representation๋งŒ ์‚ฌ์šฉ
โ€ข
๋‹ค๋ฅธ layer rep ์‚ฌ์šฉํ•˜๋ฉด ๋” ํ’๋ถ€ํ•œ ์ •๋ณด๋ฅผ ๋‹ด์„ ์ˆ˜ ์žˆ์ง€ ์•Š์„๊นŒ?
โ—ฆ
lower layer: individual features
โ—ฆ
higher layer: neighbour features