Search

2. Backgrounds and Categorization

Person
๋น„๊ณ 

2.1. Recommender Systems

โ€ข
Estimate usersโ€™ preference for item by the learnt user representation and item representation
โ€ข
user-item collaborative filtering recommendation
โ—ฆ
leverage only the user-item interaction
โ—ฆ
user representation can be explored by the sequential pattern of her/his historical interactions
โ€ข
sequential recommendation(session-based recommendation)
โ€ข
social recommendation
โ€ข
knowledge graph-based recommendation

2.2 Graph Neural Network Techniques

โ€ข
Aggregation and update operations of five typical GNN frameworks
โ€ข
GCN
โ—ฆ
๊ทธ๋ž˜ํ”„ ๋ผํ”Œ๋ผ์‹œ์•ˆ์˜ first-order eigendecomposition์„ ํ†ตํ•ด aggregation
โ€ข
Spectral GCN
1.
๊ธฐ๋ณธ์ ์ธ CNN ์—ฐ์‚ฐ์€ Graph๊ฐ€ Spatial Domain์ด ์•„๋‹ˆ๊ธฐ์— ์—ฐ์‚ฐ์ด ๋ถˆ๊ฐ€๋Šฅ
2.
์‹ ํ˜ธ์ฒ˜๋ฆฌ์—์„œ ์‚ฌ์šฉํ•˜๋Š” Convolution Theorem์„ ์‚ฌ์šฉ ( ํ‘ธ๋ฆฌ์— ๋ณ€ํ™˜ )
3.
์ด๋•Œ, graph signal์˜ ํ‘ธ๋ฆฌ์— ๋ณ€ํ™˜์€ graph์˜ Laplacian matrix๋ฅผ eigen-decompositionํ•˜๋Š” ๊ฒƒ๊ณผ ๊ฐ™์Œ
4.
3๊นŒ์ง€ ํ•˜๋ฉด ํ•™์Šต๊ฐ€๋Šฅํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๊ฐ€ ์—†๋Š”๋ฐ, ์ด๊ฑธ bruna 2014, Spectral Networks and Locally connected networks on Graph (ICLR) ์—์„œ Diagonal matrix๊ฐ€ ์›๋ž˜๋Š” eigen value์— ๋Œ€ํ•œ ํ•จ์ˆ˜์ธ๋ฐ ์ด๋ฅผ ํ•™์Šต ๊ฐ€๋Šฅํ•˜๋„๋ก ๋ฐ”๊ฟˆ
5.
ํ•˜์ง€๋งŒ, 4์˜ ๊ฒฝ์šฐ Eigen vector์— ๋Œ€ํ•œ ์˜์กด์„ฑ์ด ์žˆ์–ด์„œ ์ด์— ๋Œ€ํ•œ ์˜์กด์„ฑ์„ ์—†์• ๊ณ ์ž Approximation์„ ๋„์ž…
6.
Chebyshev Spectral CNN์œผ๋กœ ์ ‘๊ทผ (Defferrard 2016, CNNs on Graphs with Fast Localized Spectral Filtering, Nips)
7.
6์˜ ๋ฐฉ์‹์—์„œ ํŠน์ • ์กฐ๊ฑด์œผ๋กœ ๊ทผ์‚ฌํ•œ๊ฒŒ GCN
โ€ข
ChebNet์„ 1์ฐจ ๊ทผ์‚ฌ โ†’ Spatial-based
โ€ข
GraphSAGE
โ—ฆ
๊ธฐ์กด์—๋Š” ๊ณ ์ •๋œ ๋‹จ์ผ graph์— ๋Œ€ํ•ด ์˜ˆ์ธก โ†’ Transductive
โ—ฆ
์ƒˆ๋กœ์šด ๋…ธ๋“œ์— ๋Œ€ํ•ด์„œ๋„ ํ•ฉ๋ฆฌ์ ์ธ ์ถ”๋ก ์„ ํ•  ์ˆ˜ ์žˆ๊ฒŒ โ†’ Inductive
โ—ฆ
์ด์›ƒ ๋…ธ๋“œ์˜ ์œ„์ƒ์ ์ธ ๊ตฌ์กฐ๋ฅผ ํ•™์Šต โ†’ ๊ด€์ธก๋˜์ง€ ์•Š์€ ๋…ธ๋“œ์— ๋Œ€ํ•ด์„œ๋„ ์ผ๋ฐ˜ํ™” ๊ฐ€๋Šฅ
โ—ฆ
sample a fixed size of neighborhood for each node
โ—ฆ
mean/sum/max-pooling aggregator
โ€ข
GAT
โ—ฆ
attention mechanism์„ ํ†ตํ•ด ์ด์›ƒ๋…ธ๋“œ๋“ค์˜ ๊ฐ€์ค‘์น˜๋ฅผ ์ฐจ๋ณ„ํ™”
โ€ข
GGNN (Gated Graph Neural Networks)
โ—ฆ
GRU for updating
โ€ข
HGNN
โ—ฆ
hypergraph neural network

2.3 Why Graph Neural Network for Recommendation

โ€ข
๊ทธ๋ž˜ํ”„ ๋ฐ์ดํ„ฐ์˜ representation learning์— ์žˆ์–ด์„œ์˜ ์ ํ•ฉ์„ฑ
โ€ข
non-sequential data์—์„œ์˜ node representation, sequential data์—์„œ์˜ graph representation(sequence representation) ๋ชจ๋‘ ๊ฐ€๋Šฅ
โ€ข
Flexibility to incorporate additional information
โ—ฆ
Ex. social network graph๋ฅผ user-item bipartite graph์™€ ํ•จ๊ป˜ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Œ
โ€ข
GNN can encode collaborative signal of user-item interactions
โ€ข
Flexible to model multi-hop connectivity

2.4 Categories of Graph Neural Network-based Recommendation