Dynamic Network Link Prediction by Learning Effective Subgraphs using CNN-LSTM Kalyani Selvarajah School of Computer Science University of Windsor Windsor, ON, Canada selva111@uwindsor.ca Kumaran Ragunathan School of Computer Science University of Windsor Windsor, ON, Canada ragunat@uwindsor.ca Ziad Kobti School of Computer Science University of Windsor Windsor, ON, Canada kobti@uwindsor.ca Mehdi Kargar Ted Rogers School of Management Ryerson University Toronto, ON, Canada kargar@ryerson.ca Abstract—Predicting the future link between nodes is a sig- nificant problem in social network analysis, known as Link Prediction (LP). Recently, dynamic network link prediction has attracted many researchers due to its valuable real-world applications. However, most methods fail to perform satisfying prediction accuracy in various types of networks because the dynamic LP in evolving networks is struggling with spatial and nonlinear transitional patterns. Besides this, existing methods mostly involve the whole network and target link for the LP process. It leads to high computational costs. This paper aims to address these issues by proposing a novel framework named DLP-LES using deep learning methods. DLP-LES uses common neighbors based subgraph of a target link and learns the transitional pattern of it for a given dynamic network. We extract a set of heuristic features of the evolving subgraph to gather additional information about the target link. In this way, we avoid examining the entire network. Additionally, our model introduces new mechanisms to reduce computational costs. DLP- LES generates a lookup table to keep the required information of links of the network and uses a hashing method to store and fetch link information. We propose an algorithm to construct feature matrices of the evolving subgraph to learn transitional link patterns. Our model transforms the dynamic link prediction to a video classification problem, and uses Convolutional Neural Networks with Long Short-Term Memory neural networks. To verify the effectiveness of DLP-LES, extensive experiments are carried out on five real-world dynamic networks. We compare those results against four network embedding methods and basic heuristic methods. Index Terms—Social Networks, Dynamic Networks, Link Pre- dictions, Subgraphs I. I NTRODUCTION Dynamic network analysis has become an important re- search problem in recent years because it resembles the evolving nature of real-world networks. It has taken a great deal of attention from various fields, including social science [1], economics [2], and biology [3]. Dynamic networks evolve over time, and nodes and links may appear or disappear as time goes by. One of the primary areas of research in dynamic networks is temporal link prediction, which attempts to predict the links in the future using the transformation of a sequence of networks. LP has several applications including friend recommendation [4], classify the behavior and motion of people [5], and disease gene prediction [6]. Numerous studies have been performed in a static network setting, which considers a single snapshot of a network at time Figure 1. The representation of a dynamic network G with a series of snapshots from time 1 to t as a input and a snapshot at time t +1 as a output t and is used to determine new links in time t ′ (>t). Simple heuristic methods, often based on topological properties of the network, such as common neighbors [7], Adamic-Adar [4] and Katz [8] or a combination of such heuristics are well-defined for static networks. Link prediction in a dynamic network is a challenging and complex process. It has a completely new dimension of analysis because the history of network evolution provides more information to detect potential or future links. The dynamic network settings can be generally formulated as the sequence of network snapshots, as shown in Figure 1, where the behavior of each snapshot can be described as a static network at a time. To deal with dynamic network link prediction, various methods have been proposed in the literature [9]–[12]. These methods include network embedding techniques such as DeepWalk [13], LINE [14] and Node2Vec [12] and deep learning techniques [9]–[11]. The approach in [15] and [16] have explored the usage of heuristic methods in the dynamic network link prediction. Most of the existing approaches in both static and dynamic settings focused only on the target nodes, source and des- tination of the link and entire network for the prediction. However, the target nodes and their neighbor nodes play a high impact on link prediction, and analyzing the portion of the whole network reduce time complexity. Recent ground- breaking methods in static networks, WLMN [17] and SEAL [18] proposed neural network approaches to automate the selection of best heuristic for a given network, and introduced subgraph extraction methods, based on neighbor nodes, of the target links for the prediction. However, PLACN [19] claimed that subgraphs by common neighbor nodes of target link have 978-1-7281-6926-2/20/$31.00 ©2020 IEEE