[Paper Review] GC-LSTM: Graph Convolution Embedded LSTM for Dynamic Link Prediction

Jinyin Chen, Wang, Xueke+1 more|arXiv (Cornell University)|Dec 11, 2018

Complex Network Analysis TechniquesPhysics and Astronomy53 references85 citations

TL;DR

GC-LSTM embeds graph convolution into LSTM to learn spatio-temporal features for dynamic network link prediction, predicting added, removed, or preserved links. It outperforms state-of-the-art baselines on real datasets.

ABSTRACT

Dynamic link prediction is a research hot in complex networks area, especially for its wide applications in biology, social network, economy and industry. Compared with static link prediction, dynamic one is much more difficult since network structure evolves over time. Currently most researches focus on static link prediction which cannot achieve expected performance in dynamic network. Aiming at low AUC, high Error Rate, add/remove link prediction difficulty, we propose GC-LSTM, a Graph Convolution Network (GC) embedded Long Short Term Memory network (LTSM), for end-to-end dynamic link prediction. To the best of our knowledge, it is the first time that GCN embedded LSTM is put forward for link prediction of dynamic networks. GCN in this new deep model is capable of node structure learning of network snapshot for each time slide, while LSTM is responsible for temporal feature learning for network snapshot. Besides, current dynamic link prediction method can only handle removed links, GC-LSTM can predict both added or removed link at the same time. Extensive experiments are carried out to testify its performance in aspects of prediction accuracy, Error Rate, add/remove link prediction and key link prediction. The results prove that GC-LSTM outperforms current state-of-art method.

Motivation & Objective

• Motivate dynamic network link prediction (DNLP) and its broad applications.
• Propose an end-to-end model that jointly learns structural and temporal features for DNLP.
• Enable prediction of both new and disappearing links, not just additions.
• Evaluate GC-LSTM against existing baselines on multiple real-world dynamic networks.

Proposed method

• Represent each snapshot as an adjacency matrix and use a sequence of past adjacency matrices as input.
• Embed Graph Convolution into the LSTM by applying GCN to both the hidden state and the cell state (two GCNs per LSTM cell).
• Use a two-part encoder-decoder framework where GC-LSTM acts as the encoder and a fully connected decoder maps hidden states to link probabilities.
• Utilize a Chebyshev polynomial-based graph convolution to approximate spectral filters on directed graphs (K-th order with K=3).
• Train the model with a loss combining L2 reconstruction error on the adjacency matrix and a regularization term, optimized with Adam.

Experimental results

Research questions

• RQ1 How can structural information from each network snapshot be integrated with temporal evolution for DNLP?
• RQ2 Can embedding GCNs into LSTM cells improve prediction of both added and removed links in dynamic networks?
• RQ3 What is the impact of the graph convolution order (K) on predictive performance and efficiency?
• RQ4 How does GC-LSTM compare to existing DNLP methods on real-world datasets across multiple metrics?

Key findings

• GC-LSTM effectively learns both spatial structure and temporal evolution for dynamic networks.
• The model predicts added, removed, and constant links, enabling full dynamic network evolution prediction.
• Experiments show GC-LSTM outperforms state-of-the-art baselines on several real-world datasets across standard metrics.
• GC-LSTM uses K=3 Chebyshev-based graph convolution and a 256–512 dimensional hidden state depending on dataset, balancing performance and efficiency.