Domain-Informed Negative Sampling Strategies for Dynamic Graph Embedding in Meme Stock-Related Social Networks

TL;DR

This work targets dynamic graph embedding for meme stock–related social networks and identifies that standard negative sampling strategies fail to capture domain-specific interaction patterns. The authors introduce domain-informed negative sampling (DINS), combining multiple NSS tailored to network dynamics such as timing of repeated interactions, role reversals, and loop formation, plus positive enhancement to balance data. Through extensive experiments on Reddit datasets for GME, AMC, and BB with three DGE models (TGNs, DyGFormer, GraphMixer), DINS improves predictive performance and provides a more comprehensive evaluation framework. The findings highlight the practical value of incorporating domain knowledge into NSS design for dynamic graphs and its potential to better understand meme stock–driven market dynamics.

Abstract

Social network platforms like Reddit are increasingly impacting real-world economics. Meme stocks are a recent phenomena where price movements are driven by retail investors organizing themselves via social networks. To study the impact of social networks on meme stocks, the first step is to analyze these networks. Going forward, predicting meme stocks' returns would require to predict dynamic interactions first. This is different from conventional link prediction, frequently applied in e.g. recommendation systems. For this task, it is essential to predict more complex interaction dynamics, such as the exact timing. These are crucial for linking the network to meme stock price movements. Dynamic graph embedding (DGE) has recently emerged as a promising approach for modeling dynamic graph-structured data. However, current negative sampling strategies, an important component of DGE, are designed for conventional dynamic link prediction and do not capture the specific patterns present in meme stock-related social networks. This limits the training and evaluation of DGE models in such social networks. To overcome this drawback, we propose novel negative sampling strategies based on the analysis of real meme stock-related social networks and financial knowledge. Our experiments show that the proposed negative sampling strategies can better evaluate and train DGE models targeted at meme stock-related social networks compared to existing baselines.

Figures (4)

• Figure 1: Visualization of proposed NSSs. Taking a dynamic network with 8 interactions as an example. The batch size is 2, meaning that the interactions are processed in groups of two. The showed negative samples are generated according to positive sample $(u,v,t_3)$.
• Figure 2: Evaluation results of DyGFormer trained with three different NSSs. The blue bar, green bar and orange bar (from left to right) show the results of model trained with random NSS, historical NSS and DINS (proposed).
• Figure 3: Evaluation results of GraphMixer trained with three different NSSs. The blue bar, green bar and orange bar (from left to right) show the results of model trained with random NSS, historical NSS and DINS (proposed).
• Figure 4: Evaluation results of TGNs trained with three different NSSs. The blue bar, green bar and orange bar (from left to right) show the results of model trained with random NSS, historical NSS and DINS (proposed).

Theorems & Definitions (4)

• Definition 1: Continuous dynamic graph
• Definition 2: Dynamic Graph Embedding
• Definition 3: Dynamic Link Prediction
• Definition 4: Negative Sampling Strategy