Fine-Grained Double-View Link Prediction Within the Dynamic Interaction Network

Trend prediction in financial trading markets is crucial in decision-making support, normally realized via mining trading patterns or analyzing technique indicators. However, cryptocurrency decentralized exchange (DEX) markets are growing rapidly recently and draw much attention. Different from the centralized market, DEX allows users to observe the whole market data at the order-level, where identity information associated with orders is also available. Moreover, trading pairs in DEX are often manipulated by bots, and user ordering behavior may be affected by these algorithm-controlled bots. These provide us with a novel human-bot game scenario to investigate the predictability of human ordering behavior. In this paper, we model the trading market of the cryptocurrency DEX at each timestamp as a user-order bipartite graph and propose a double-view dynamic network for fine-grained user ordering prediction in a human-bot mixed DEX market. Given the ordering information of the humans and bots, we learn current latent representations of humans and bots using graph neural networks and recurrent neural networks respectively for user ordering prediction. We conduct experiments on real cryptocurrency data from Binance Dex. The experimental results of link prediction demonstrate that our proposed method outperforms the state-of-the-art models.