Network Embedding Overview and Recent Research Directions from CIKM 2017

The 26th ACM International Conference on Information and Knowledge Management (CIKM 2017) featured a strong focus on network embedding, with 22 of the 171 long papers related to graph mining and 11 specifically on network representation learning.

Network embedding projects nodes and communities of a graph into a low‑dimensional vector space to support tasks such as node classification, link prediction, community detection, and visualization. The core assumption is that closer nodes in the graph should have closer embedding vectors, and the loss function typically penalizes distance discrepancies.

Algorithmic approaches are grouped into three main families: (1) matrix‑based spectral methods that preserve pairwise similarity; (2) random‑walk based methods such as DeepWalk and Node2Vec that generate node sequences and learn embeddings via skip‑gram models; (3) deep‑learning models including SDNE (Structural Deep Network Embedding) which use auto‑encoders, and Graph Convolutional Networks (GCN) that incorporate node attributes and neighbor information.

Recent research directions highlighted at CIKM include: dynamic and attributed network embedding that jointly models evolving graph structure and node features (e.g., DANE), meta‑path based heterogeneous network embedding (HIN2Vec), attention‑driven multi‑view representation learning, community embedding that simultaneously learns node and community vectors, and name disambiguation in anonymized graphs using only relational information.

Practical application scenarios discussed range from two‑step recommendation in attributed signed networks, real‑time risk identification with incremental updates, to online fraud detection and heterogeneous information retrieval. Each direction is illustrated with representative papers and visual diagrams of model architectures and loss formulations.