Recommender System
Graph Neural Networks in Recommender Systems: A Survey, arxiv 2020
For e-commerce and social media platforms, a recommender system is important. Motivation for GNN on RS: (1) RS data is essentially graphs:
Relevant information includes social relationships, bipartite user-item interaction graphs, item transitions, which can all be represented well as a graph structure. (2) GNN is good on graphs.
General Recommendation
static preference. model user-item compatibility by implicit (click) or explicit (rating)
Can utilize side knowledge:
- social network → influence modeling (friends); preference integration (user-item; social)
- knowledge graph → graph simplification (kg too complicate); multi-relation propagation; user integration
Sequential Recommendation
Transition pattern for next item recommendation.
The sequential recommendation assumes the most recent click = dynamic preference → uses Markov chain to model.
User implicit/explicit feedback on items as bipartite graph connectivity matrix of m users ×n items.
Can be enhanced with social networks.
Future directions
- Efficient GNNs for Heterogeneous Graphs
- Multi-graph Information Integration
- Scalability
- Consider time for sequential recommendations
Streaming Data
- There has been little focus on graph neural networks for dynamic networks.
- Not concerning spatial-temporal graphs (only feature change, no structure change)
- Many names: dynamic networks, temporal networks, evolutionary networks, or time-varying networks.
(Streaming data??) - Encoder-Decoder framework.
Section II: dynamic network; Section III: a survey on encoders; Section IV: for prediction — decoder, loss, metrics.
Section II: dynamic network
Add start and end time to every node and edge of graph
G = (V, E) = ({(v, t_s, t_e)}; {u, v, t_s, t_e})
## edge existence time should be constraint by node existence time
—[A] Dynamic Network Representation —
Static Graph → Edge weighted: label edge by active time
→ Discrete: marked by discrete time intervals, e.g. snapshots
→ Continous: the most complex
Discrete Representation — DG = {(G^1, G^2, G^3… … G^T)}
Continous Representations —
(i) event-based: focused on link/edge duration
EB = {(u_i, v_i, t_i, Δ_i); i = 1,2,3…}
(ii) contact sequence: link duration is short or unimportant
CS = {(u_i, v_i, t_i); i = 1,2,3…}
(iii) graph stream: used to circumvent hardware limit?
GS = {(u_i , v_i , t_i , δ_i); i=1,2,3…}, δ_i=add (1), or remove (-1)
###the most general G = (V, E) = ({(v, t_s, t_e)}; {u, v, t_s, t_e}) not discussed??
—[B] Link duration spectrum —
E.g. email (instant) < proximity (of person) < employment network < internet < citation (static)
— [C] Node dynamic —
Stacie; Dynamic (special case: Growing)
Machine learning works for supervised classifying streaming: changes in the data distribution over time, i.e. concept drift.
Not much work on semi-supervised, unsupervised learning