STAMP 모델의 구조와 거의 비슷하게 진행된다. 사실상 기존에 있는 Gated-GNN과 STAMP 모델을 합친거라고 볼수 있다. Gated-GNN을 통해서 Session 안에서 클릭된 아이템 간의 transition을 모델링한다. 여기서는 As를 여러 가지로 정할 수 있다. 여기서 item node의 transition이 잘 모델링 되어있는지에 대해서 고민이 좀 더 필요하다. 결국에는 directed graph로는 모델링이 안된 것이라고 볼 수도 있을 것 같다. 왜냐면 connection matrix가 outgoing edges, incoming edges로 나눠져있지만 이걸 계산 과정에서 어떻게 따로 보는지에 대해서 납득이 안되기 때문이다. 각 단계에서 계산되어지는 값들에 대한 이해가 필요하다. 그리고 STAMP 모델 처럼 session의 global preference와 current interest를 나누어서 계산하고, current interest는 session에서 마지막으로 클릭된 아이템의 embedding으로 계산된다. 그리고 global preference는 attention network를 거쳐서 각 아이템에 대한 가중치를 계산한 후에 그 weighted sum으로 계산된다. 이 모델의 구조가 optimal은 아닐 것 같다. 여기서 더 개선시킬 수 있는 방법이 있을 것 같다. attention network를 대체할 수 있는 구조가 있거나, current interest를 last clicked item embedding으로 단순하게 보는 것 이상으로 무언가를 만들 수 있지 않을까?
SR-GNN model
Learning node vector with GRU cell
update function (GRU cell)
Session representation with current interest and global preference
Generating session embedding
current interest = list click item embedding
global session preference
attention mechanism
linear transformation
Making Recommendation and training
Performance Comparison
Datasets
Yoochoose, Diginetica
Reference
Wu, Shu, et al. "Session-based recommendation with graph neural networks." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019. 'PAPER'
Research Area
Session-based Recommendation
Key points
이 논문은 Gated-GNN을 사용해서 Node embedding을 학습시킨다.
Model ideas
STAMP 모델의 구조와 거의 비슷하게 진행된다. 사실상 기존에 있는 Gated-GNN과 STAMP 모델을 합친거라고 볼수 있다. Gated-GNN을 통해서 Session 안에서 클릭된 아이템 간의 transition을 모델링한다. 여기서는 As를 여러 가지로 정할 수 있다. 여기서 item node의 transition이 잘 모델링 되어있는지에 대해서 고민이 좀 더 필요하다. 결국에는 directed graph로는 모델링이 안된 것이라고 볼 수도 있을 것 같다. 왜냐면 connection matrix가 outgoing edges, incoming edges로 나눠져있지만 이걸 계산 과정에서 어떻게 따로 보는지에 대해서 납득이 안되기 때문이다. 각 단계에서 계산되어지는 값들에 대한 이해가 필요하다. 그리고 STAMP 모델 처럼 session의 global preference와 current interest를 나누어서 계산하고, current interest는 session에서 마지막으로 클릭된 아이템의 embedding으로 계산된다. 그리고 global preference는 attention network를 거쳐서 각 아이템에 대한 가중치를 계산한 후에 그 weighted sum으로 계산된다. 이 모델의 구조가 optimal은 아닐 것 같다. 여기서 더 개선시킬 수 있는 방법이 있을 것 같다. attention network를 대체할 수 있는 구조가 있거나, current interest를 last clicked item embedding으로 단순하게 보는 것 이상으로 무언가를 만들 수 있지 않을까?
SR-GNN model
Learning node vector with GRU cell
Session representation with current interest and global preference
Generating session embedding
current interest = list click item embedding
global session preference
linear transformation
Making Recommendation and training
Performance Comparison
Datasets
Yoochoose, Diginetica
Reference
Wu, Shu, et al. "Session-based recommendation with graph neural networks." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. 2019. 'PAPER'