Predicting Blockchain Application Performance with Machine Learning Techniques

Abstract

Blockchain technology is a distributed ledger designed to record all transactions within its network, characterized by its decentralized nature, resistance to tampering, and attributes such as consistency, anonymity, and traceability. However, evaluating blockchain applications' performance can be complex due to their intricate and distributed infrastructure. This research employs machine learning model-based methods to predict blockchain systems' performance using predetermined configuration parameters. The data used in this study is derived from a blockchain simulator, generating blockchain data to facilitate performance predictions. The simulation process involves using simulated data and configuration settings for each run, including parameters such as the number of nodes, the number of miners, consensus algorithm, maximum block size, and transaction quantities, among others. Output metrics such as the total number of blocks, transaction rate, block propagation time, and latency are utilized to assess network performance. The simulator was run 184 times with various configurations. Our findings indicate that the Random Forest model outperformed other models used in the experiments, achieving the highest R² scores for multiple metrics, such as 0.987 for total number of transactions and 0.765 for average block propagation time, while also demonstrating lower RMSE values, indicating more accurate predictions.