Influence Maximization in Temporal Social Networks with a Cold-Start Problem: A Supervised Approach

TL;DR

This work tackles Influence Maximization in temporal social networks under a cold-start scenario. It defines Influence Propagation Paths (IPPs) and uses Motif-Based Filtering to generate ground-truth IPPs, paired with a tensorized Temporal Graph Network (TGN) for multi-relational graphs to predict seeds efficiently. A novel cold-start augmentation serializes IPPs and employs a prefix-tree retrieval to connect sparse neighborhoods, improving diffusion, especially in early propagation. Offline evaluations and online A/B tests in a real team-based game demonstrate consistent improvements in network growth, with significant gains for cold-start cases and substantial training-speedups over traditional TGNN approaches.

Abstract

Influence Maximization (IM) in temporal graphs focuses on identifying influential "seeds" that are pivotal for maximizing network expansion. We advocate defining these seeds through Influence Propagation Paths (IPPs), which is essential for scaling up the network. Our focus lies in efficiently labeling IPPs and accurately predicting these seeds, while addressing the often-overlooked cold-start issue prevalent in temporal networks. Our strategy introduces a motif-based labeling method and a tensorized Temporal Graph Network (TGN) tailored for multi-relational temporal graphs, bolstering prediction accuracy and computational efficiency. Moreover, we augment cold-start nodes with new neighbors from historical data sharing similar IPPs. The recommendation system within an online team-based gaming environment presents subtle impact on the social network, forming multi-relational (i.e., weak and strong) temporal graphs for our empirical IM study. We conduct offline experiments to assess prediction accuracy and model training efficiency, complemented by online A/B testing to validate practical network growth and the effectiveness in addressing the cold-start issue.

Figures (6)

• Figure 1: In the graph constructed from deduplicated edges derived from daily invitations and adoptions, the proportion of nodes with only one neighbor over the total number of nodes (y-axis) is presented over a span of $30$ days (x-axis).
• Figure 2: Pipeline for predicting active members. Weak and strong graphs are constructed using exposure edges and invitation and adoption edges derived from in-game teammate recommendations, respectively. These graphs are then input into a Tensorized TGN to generate node scores. During training, node scores with labels from IPP findings are used to calculate the training loss. In the inference phase, nodes with top scores are active members.
• Figure 3: Cold-start solution: (A) IPP Serialization: Serialized strings are generated for each node in an IPP and then concatenate into an IStr. (B) Neighbor Retrieval: The IStrs of all IPPs are inserted into a prefix tree and positioned by pre-order traversal (PTT), ensuring that similar strings are placed adjacently. Cold-start edges are established between the end node $v_2$ of an IPP and other nodes belonging to the most similar IStrs in the PTT. (C) Demonstration of neighbor retrieval.
• Figure 4: Proportion of network scale over $6$ days.
• Figure 5: Investigation of degree v.s. hop, where x,y position refers to $\{1,2,3\}\times\{1,2,6\}$. Each figure presents the proportion of network scale over $6$ days.

Theorems & Definitions (3)

• Definition 1
• Definition 2
• Definition 3